AI-generated content is becoming increasingly difficult to distinguish from human-created work. A convincing photograph might never have been taken, a video might show something that never happened, and a well-written article might have been produced by AI. So what clues should we look for, and can we really trust AI detection tools? Explore eight perspectives on recognising AI-generated content, verifying sources and developing the critical thinking skills needed to navigate today’s digital world.
BEFORE YOU READ...
Would you recognise an AI-generated image if it appeared in your social media feed?
Think about the last surprising image, video or story you encountered online. Did you question whether it was real? What clues would you look for, and how could you check whether the source was trustworthy?
Explore Different Perspectives
Discover eight perspectives on identifying AI-generated content, from unusual visual details and writing patterns to digital watermarking, source verification and media literacy.
Remember: A clue is not proof. Unusual details, writing patterns and AI detection tools can raise questions about a piece of content, but none can establish its origin with certainty. When authenticity matters, investigate the original source and look for independent verification.
Ricard Pena Cerdan — Can we still trust what we see online? The challenge of spotting AI content
That is a challenge that is going to increase its difficulty progressively in the incoming years. As the generative AI will be trained more and better and the models will improve, the small signs of AI generated content are going to become almost impossible to distinguish. So I would recommend, as the first measure to be sure of the content’s origins, to not trust any content we can find online without a proper human signature. And even if that firm is there, maybe it would be healthy to still doubt its source. We are, sadly, entering a time in human history where a text or photography is not a proof of facts anymore. For the most empirical approach I would point out some signs that have helped me make a distinction between AI generated content and the human one. Text-wise, language models usually go for short enunciative sentences with not much rhetorical figures and self-explanatory points of view. Either if they are supporting or criticizing any point, they won’t mix different povs. Talking about images or videos, usually they have some problems filling pixels in complex structures, so if the observer is able to magnify the image it would be easy to spot some inconsistency. As I say, most of these issues are being fixed by many models right now, so it’s a problem every part of society will need to attend to at some point with no clear outcome.
Ricard Peña Cerdan, Denmark, a journalist, AI Enthusiast
Læs på dansk (DK)
Det er en udfordring, som gradvist vil blive vanskeligere i de kommende år. Efterhånden som generativ AI bliver trænet mere og bedre, og modellerne forbedres, vil de små tegn på AI-genereret indhold blive næsten umulige at skelne. Derfor vil jeg som en første forholdsregel for at kunne være sikker på indholdets oprindelse anbefale, at man ikke stoler på noget indhold, man finder online, hvis det ikke har en tydelig menneskelig afsender. Og selv hvis der er en sådan afsender, kan det stadig være sundt at stille spørgsmål ved kilden. Vi er desværre på vej ind i en periode af menneskets historie, hvor en tekst eller et fotografi ikke længere er et bevis på faktiske forhold. Hvis jeg skal anlægge en mere empirisk tilgang, vil jeg fremhæve nogle tegn, der har hjulpet mig med at skelne mellem AI-genereret og menneskeskabt indhold. Når det gælder tekst, har sprogmodeller som regel en tendens til at bruge korte, konstaterende sætninger uden mange retoriske virkemidler og med synspunkter, der forklarer sig selv. Uanset om de støtter eller kritiserer et bestemt synspunkt, blander de normalt ikke forskellige perspektiver. Når det gælder billeder eller videoer, har modellerne ofte problemer med at gengive pixels korrekt i komplekse strukturer. Hvis man kan forstørre billedet, vil det derfor være lettere at opdage uoverensstemmelser. Som sagt er mange modeller allerede i gang med at løse de fleste af disse problemer, så det er en udfordring, som alle dele af samfundet på et tidspunkt bliver nødt til at forholde sig til, uden at der endnu er nogen klar løsning.Ricard Peña Cerdan, Danmark, journalist, AI-entusiast
Auf Deutsch lesen (DE)
Das ist eine Herausforderung, die in den kommenden Jahren zunehmend schwieriger werden wird. Je mehr und besser generative KI trainiert wird und je weiter sich die Modelle verbessern, desto schwieriger wird es, die kleinen Anzeichen für KI-generierte Inhalte zu erkennen. Deshalb würde ich als erste Maßnahme zur Überprüfung der Herkunft von Inhalten empfehlen, keinem Inhalt im Internet zu vertrauen, der nicht eindeutig einer menschlichen Urheberin oder einem menschlichen Urheber zugeordnet werden kann. Und selbst wenn eine solche Zuordnung vorhanden ist, kann es sinnvoll sein, die Quelle weiterhin zu hinterfragen. Leider treten wir in eine Phase der Menschheitsgeschichte ein, in der ein Text oder ein Foto nicht mehr als Beweis für Tatsachen gelten kann. Bei einer möglichst empirischen Herangehensweise würde ich auf einige Merkmale hinweisen, die mir dabei geholfen haben, KI-generierte von menschlich erstellten Inhalten zu unterscheiden. Bei Texten verwenden Sprachmodelle gewöhnlich kurze, feststellende Sätze mit wenigen rhetorischen Stilmitteln und selbsterklärenden Standpunkten. Unabhängig davon, ob sie eine Position unterstützen oder kritisieren, vermischen sie normalerweise keine unterschiedlichen Perspektiven. Bei Bildern oder Videos haben die Modelle häufig Schwierigkeiten, Pixel in komplexen Strukturen korrekt darzustellen. Wenn man das Bild vergrößern kann, lassen sich daher manche Unstimmigkeiten leichter erkennen. Wie gesagt, viele Modelle beheben die meisten dieser Probleme bereits, sodass sich irgendwann alle Bereiche der Gesellschaft mit dieser Herausforderung auseinandersetzen müssen, ohne dass es bislang eine eindeutige Lösung gibt.Ricard Peña Cerdan, Dänemark, Journalist, KI-Enthusiast
Leer en español (ES)
Ese es un desafío que será progresivamente más difícil de afrontar en los próximos años. A medida que la IA generativa se entrene más y mejor y los modelos vayan mejorando, las pequeñas señales que permiten identificar el contenido generado por IA serán casi imposibles de distinguir. Por eso, como primera medida para asegurarnos del origen de un contenido, recomendaría no confiar en nada de lo que encontremos en internet si no cuenta con una autoría humana claramente identificable. E incluso cuando exista esa firma, quizá sea saludable seguir cuestionando la fuente. Lamentablemente, estamos entrando en una etapa de la historia de la humanidad en la que un texto o una fotografía ya no constituyen una prueba de los hechos. Desde un enfoque más empírico, señalaría algunas características que me han ayudado a distinguir entre el contenido generado por IA y el creado por personas. En cuanto a los textos, los modelos de lenguaje suelen utilizar frases enunciativas cortas, con pocas figuras retóricas y puntos de vista que se explican por sí solos. Tanto si apoyan como si critican una postura, normalmente no mezclan diferentes perspectivas. En el caso de las imágenes o los vídeos, suelen tener problemas para representar correctamente los píxeles en estructuras complejas, por lo que, si el observador puede ampliar la imagen, le resultará más fácil detectar alguna incoherencia. Como decía, muchos modelos ya están solucionando la mayoría de estos problemas, así que se trata de un desafío al que todos los sectores de la sociedad tendrán que enfrentarse en algún momento, sin que exista todavía una solución clara.Ricard Peña Cerdan, Dinamarca, periodista, entusiasta de la IA
Прочетете на български (BG)
Това е предизвикателство, което постепенно ще става все по-трудно през следващите години. С напредването на обучението на генеративния ИИ и усъвършенстването на моделите малките признаци, по които може да се разпознае съдържание, генерирано от ИИ, ще стават почти невъзможни за откриване. Затова като първа мярка за установяване на произхода на дадено съдържание бих препоръчал да не се доверяваме на нищо, което намираме онлайн, ако то няма ясно посочен автор, който е реален човек. И дори когато такъв автор е посочен, може би е разумно все пак да поставяме източника под съмнение. За съжаление навлизаме в период от човешката история, в който текстът или фотографията вече не могат да служат като доказателство за факти. Ако подходим по-емпирично, бих посочил някои признаци, които са ми помогнали да различавам съдържанието, генерирано от ИИ, от това, създадено от хора. Що се отнася до текстовете, езиковите модели обикновено използват кратки съобщителни изречения, с малко реторични фигури и със самообясняващи се гледни точки. Независимо дали подкрепят или критикуват дадена позиция, те обикновено не смесват различни гледни точки. При изображенията и видеоклиповете моделите често срещат затруднения при правилното изобразяване на пикселите в сложни структури. Затова, ако наблюдателят може да увеличи изображението, ще му бъде по-лесно да забележи някои несъответствия. Както вече казах, много модели в момента отстраняват повечето от тези проблеми, така че това е предизвикателство, с което в даден момент всички части на обществото ще трябва да се справят, без засега да има ясно решение.Ricard Peña Cerdan, Дания, журналист, ентусиаст в областта на ИИ
Nicolai Christensen Steensgaard — When AI becomes impossible to spot: Why trustworthy sources matter
This distinction has become harder as AI tools have improved during the past few years, and it will likely only become even harder in the future. As of right now text is almost impossible to detect for being AI made just by looking at it. Image and video content have also improved significantly in recent years, but video still looks to be the most challenging to generate with AI. Especially video content including humans or animals can be more easily detected due to all the mistakes that can occur with these categories. However, to answer the question “How can we distinguish”, the true answer probably is that we can’t. Especially when video generation improves further. We can hope that AI generated content will be marked as AI generated, but it will definitely be an easy way to spread misinformation. Therefore, the only thing we can really do is to try seeking trustworthy sources and not believe everything we see on a random video on the internet.
Nicolai Christensen Steensgaard, Denmark, AI Expert
Læs på dansk (DK)
Det er blevet sværere at skelne mellem de to typer indhold, efterhånden som AI-værktøjerne er blevet bedre i løbet af de seneste år, og det vil sandsynligvis blive endnu vanskeligere i fremtiden. På nuværende tidspunkt er det næsten umuligt at afgøre, om en tekst er skabt af AI, blot ved at læse den. Billed- og videoindhold er også blevet væsentligt bedre i de senere år, men video ser stadig ud til at være det mest udfordrende at generere med AI. Især videoer med mennesker eller dyr kan lettere identificeres som AI-genererede på grund af de mange fejl, der kan opstå i disse kategorier. Men hvis vi skal besvare spørgsmålet »Hvordan kan vi skelne mellem dem?«, er det egentlige svar sandsynligvis, at det kan vi ikke. Især ikke når videogenerering bliver endnu bedre. Vi kan håbe, at AI-genereret indhold bliver mærket som AI-genereret, men det vil helt sikkert også være en nem måde at sprede misinformation på. Derfor er det eneste, vi reelt kan gøre, at forsøge at finde troværdige kilder og ikke tro på alt, hvad vi ser i en tilfældig video på internettet.Nicolai Christensen Steensgaard, Danmark, AI-ekspert
Auf Deutsch lesen (DE)
Diese Unterscheidung ist in den vergangenen Jahren immer schwieriger geworden, da sich KI-Werkzeuge stetig verbessert haben, und in Zukunft wird sie wahrscheinlich noch schwieriger werden. Derzeit ist es nahezu unmöglich, allein durch das Lesen eines Textes zu erkennen, ob er von KI erstellt wurde. Auch Bild- und Videoinhalte haben sich in den letzten Jahren erheblich verbessert, allerdings scheint die Erstellung von Videos mit KI weiterhin die größte Herausforderung zu sein. Besonders bei Videos mit Menschen oder Tieren lässt sich KI-generierter Inhalt aufgrund der zahlreichen Fehler, die in diesen Bereichen auftreten können, noch vergleichsweise leicht erkennen. Um jedoch die Frage „Wie können wir den Unterschied erkennen?“ zu beantworten: Die ehrliche Antwort lautet wahrscheinlich, dass wir es nicht können. Vor allem dann nicht, wenn sich die Videogenerierung weiter verbessert. Wir können hoffen, dass KI-generierte Inhalte als solche gekennzeichnet werden, doch sie werden auf jeden Fall eine einfache Möglichkeit bieten, Fehlinformationen zu verbreiten. Deshalb können wir letztlich nur versuchen, vertrauenswürdige Quellen zu finden und nicht alles zu glauben, was wir in irgendeinem Video im Internet sehen.Nicolai Christensen Steensgaard, Dänemark, KI-Experte
Leer en español (ES)
Esta distinción se ha vuelto más difícil a medida que las herramientas de IA han mejorado durante los últimos años, y probablemente será todavía más complicada en el futuro. Actualmente, es casi imposible detectar que un texto ha sido generado por IA simplemente leyéndolo. El contenido en imágenes y vídeos también ha mejorado considerablemente en los últimos años, aunque los vídeos siguen pareciendo lo más difícil de generar mediante IA. Especialmente los vídeos en los que aparecen personas o animales pueden identificarse con mayor facilidad debido a los numerosos errores que pueden producirse en estas categorías. Sin embargo, para responder a la pregunta «¿Cómo podemos distinguirlos?», probablemente la respuesta sea que no podemos. Sobre todo cuando la generación de vídeos siga mejorando. Podemos esperar que el contenido generado por IA se identifique como tal, pero sin duda también será una forma sencilla de difundir desinformación. Por lo tanto, lo único que realmente podemos hacer es intentar buscar fuentes fiables y no creernos todo lo que vemos en cualquier vídeo que encontramos por internet.Nicolai Christensen Steensgaard, Dinamarca, experto en IA
Прочетете на български (BG)
Разграничаването между двата вида съдържание става все по-трудно, тъй като инструментите с ИИ се усъвършенстваха през последните няколко години, а в бъдеще вероятно ще бъде още по-трудно. В момента е почти невъзможно да се установи дали даден текст е създаден от ИИ само чрез прочитането му. Изображенията и видеосъдържанието също се подобриха значително през последните години, но генерирането на видео все още изглежда най-голямото предизвикателство за ИИ. Особено видеоклиповете, в които присъстват хора или животни, могат да бъдат разпознати по-лесно като генерирани от ИИ заради многобройните грешки, които могат да възникнат при тези категории. Въпреки това, ако трябва да отговорим на въпроса „Как можем да ги различим?“, истинският отговор вероятно е, че не можем. Особено когато генерирането на видео се усъвършенства още повече. Можем да се надяваме, че съдържанието, генерирано от ИИ, ще бъде обозначавано като такова, но то със сигурност ще предоставя и лесен начин за разпространение на дезинформация. Затова единственото, което наистина можем да направим, е да се стремим да намираме надеждни източници и да не вярваме на всичко, което виждаме в случаен видеоклип в интернет.Nicolai Christensen Steensgaard, Дания, експерт по ИИ
Maialen Ayestaran — Too perfect or too uniform? Recognising the subtle clues of AI-generated content
I believe that, with experience, you start to notice certain patterns. Sometimes, content generated by artificial intelligence can have a style that is too structured, very uniform, or even slightly impersonal, and these characteristics can serve as clues that it was AI-generated. Recognizing these subtle hints can help users develop a better sense of when AI might be involved.
However, beyond these indicators, it is not always easy to distinguish AI-generated content, to be honest. For this reason, it may be most useful to rely on more sophisticated or specialized tools designed specifically to detect AI-generated material. Even with these tools, though, the task is becoming increasingly complex, and I believe that over time it will be even more difficult to differentiate between human-produced and AI-generated content. As AI technology continues to advance, the lines are likely to blur further, making critical assessment and verification even more essential.
Ayestaran Maialen, Spain, Student of Psychology, Enthusiast about AI
Læs på dansk (DK)
Jeg tror, at man med erfaring begynder at lægge mærke til bestemte mønstre. Nogle gange kan indhold, der er genereret af kunstig intelligens, have en stil, som er for struktureret, meget ensartet eller endda en smule upersonlig. Disse karakteristika kan være tegn på, at indholdet er AI-genereret. At kunne genkende disse subtile tegn kan hjælpe brugerne med at udvikle en bedre fornemmelse for, hvornår AI kan have været involveret. Men ud over disse tegn er det ærligt talt ikke altid let at skelne AI-genereret indhold fra menneskeskabt indhold. Derfor kan det være mest nyttigt at benytte mere avancerede eller specialiserede værktøjer, der er udviklet specifikt til at opdage AI-genereret materiale. Selv med disse værktøjer bliver opgaven dog stadig mere kompleks, og jeg tror, at det med tiden vil blive endnu vanskeligere at skelne mellem indhold skabt af mennesker og indhold genereret af AI. Efterhånden som AI-teknologien fortsætter med at udvikle sig, vil grænserne sandsynligvis blive endnu mere udviskede, hvilket gør kritisk vurdering og kontrol af oplysninger endnu vigtigere.Ayestaran Maialen, Spanien, psykologistuderende, AI-entusiast
Auf Deutsch lesen (DE)
Ich glaube, dass man mit zunehmender Erfahrung beginnt, bestimmte Muster zu erkennen. Manchmal haben Inhalte, die von künstlicher Intelligenz erstellt wurden, einen Stil, der zu stark strukturiert, sehr einheitlich oder sogar etwas unpersönlich wirkt. Diese Merkmale können Hinweise darauf sein, dass die Inhalte KI-generiert sind. Das Erkennen solcher subtilen Anzeichen kann Nutzerinnen und Nutzern helfen, ein besseres Gespür dafür zu entwickeln, wann KI im Spiel sein könnte. Abgesehen von diesen Hinweisen ist es jedoch ehrlich gesagt nicht immer einfach, KI-generierte Inhalte zu erkennen. Deshalb kann es besonders hilfreich sein, auf fortschrittlichere oder spezialisierte Werkzeuge zurückzugreifen, die ausdrücklich dafür entwickelt wurden, KI-generiertes Material zu identifizieren. Allerdings wird diese Aufgabe selbst mit solchen Werkzeugen zunehmend komplexer, und ich glaube, dass es mit der Zeit noch schwieriger werden wird, zwischen von Menschen erstellten und KI-generierten Inhalten zu unterscheiden. Da sich die KI-Technologie ständig weiterentwickelt, werden die Grenzen wahrscheinlich noch stärker verschwimmen, wodurch eine kritische Bewertung und Überprüfung immer wichtiger werden.Ayestaran Maialen, Spanien, Psychologiestudentin, KI-Enthusiastin
Leer en español (ES)
Creo que, con la experiencia, empiezas a identificar ciertos patrones. A veces, el contenido generado por inteligencia artificial puede tener un estilo demasiado estructurado, muy uniforme o incluso ligeramente impersonal, y estas características pueden servir como pistas de que ha sido generado por IA. Reconocer estas señales sutiles puede ayudar a los usuarios a desarrollar una mayor capacidad para detectar cuándo podría haber intervenido la IA. Sin embargo, más allá de estos indicios, sinceramente no siempre es fácil distinguir el contenido generado por IA. Por este motivo, puede resultar más útil recurrir a herramientas más avanzadas o especializadas, diseñadas específicamente para detectar material generado por inteligencia artificial. Aun así, incluso con estas herramientas, la tarea es cada vez más compleja, y creo que con el tiempo será todavía más difícil diferenciar entre el contenido producido por personas y el generado por IA. A medida que la tecnología de inteligencia artificial siga avanzando, es probable que los límites se vuelvan aún más difusos, haciendo que la evaluación crítica y la verificación sean todavía más esenciales.Ayestaran Maialen, España, estudiante de Psicología, entusiasta de la IA
Прочетете на български (BG)
Смятам, че с натрупването на опит човек започва да забелязва определени закономерности. Понякога съдържанието, генерирано от изкуствен интелект, може да има прекалено структуриран, много еднообразен или дори леко безличен стил. Тези характеристики могат да бъдат признаци, че съдържанието е създадено от ИИ. Разпознаването на подобни фини сигнали може да помогне на потребителите да развият по-добър усет за това кога е възможно да е използван ИИ. Въпреки това, ако оставим настрана тези признаци, честно казано, невинаги е лесно да се разпознае съдържание, генерирано от ИИ. Поради тази причина може да бъде най-полезно да се разчита на по-усъвършенствани или специализирани инструменти, създадени специално за откриване на материали, генерирани от изкуствен интелект. Дори с тези инструменти обаче задачата става все по-сложна и смятам, че с течение на времето ще бъде още по-трудно да се прави разлика между съдържание, създадено от хора, и съдържание, генерирано от ИИ. С продължаващото развитие на технологиите за изкуствен интелект границите вероятно ще стават още по-размити, което ще направи критичната оценка и проверката на информацията още по-важни.Ayestaran Maialen, Испания, студентка по психология, ентусиастка в областта на ИИ
Xavi — What gives AI-written text away? Looking beyond the words
It’s difficult, especially because I’m not familiar with all the styles that ChatGPT can generate. However, I do think there are certain clues that can help. For example, the vocabulary used by AI tends to be quite stable and uniform. In addition, you can look at the “depth” of the content: a class assignment or a more superficial text might raise suspicions of being AI-generated, whereas a scientific article or an academic text usually has a more precise and detailed structure. It’s not a foolproof method, but I think paying attention to consistency in vocabulary, clarity, and the apparent complexity of the content can help identify whether it has been created by AI. It’s also useful to compare it with your own knowledge of the topic; if something seems too “perfect” or overly neutral, it could be a sign of automated generation.
Xavi, Spain, student in Psychology, AI Enthusiast
Læs på dansk (DK)
Det er svært, især fordi jeg ikke kender alle de forskellige stilarter, som ChatGPT kan generere. Jeg mener dog, at der er visse tegn, som kan hjælpe. For eksempel har det ordforråd, AI bruger, en tendens til at være ret stabilt og ensartet. Derudover kan man se på indholdets »dybde«: En skoleopgave eller en mere overfladisk tekst kan vække mistanke om, at den er AI-genereret, mens en videnskabelig artikel eller en akademisk tekst som regel har en mere præcis og detaljeret struktur. Det er ikke en ufejlbarlig metode, men jeg tror, at man ved at være opmærksom på ordforrådets ensartethed, tydeligheden og indholdets tilsyneladende kompleksitet bedre kan vurdere, om det er skabt af AI. Det er også nyttigt at sammenligne indholdet med sin egen viden om emnet. Hvis noget virker for »perfekt« eller overdrevent neutralt, kan det være et tegn på, at det er genereret automatisk.Xavi, Spanien, psykologistuderende, AI-entusiast
Auf Deutsch lesen (DE)
Das ist schwierig, vor allem weil ich nicht mit allen Schreibstilen vertraut bin, die ChatGPT erzeugen kann. Allerdings glaube ich, dass es bestimmte Hinweise gibt, die helfen können. Beispielsweise ist der von KI verwendete Wortschatz häufig recht gleichbleibend und einheitlich. Außerdem kann man auf die „Tiefe“ des Inhalts achten: Eine schulische Aufgabe oder ein eher oberflächlicher Text könnte den Verdacht wecken, dass er KI-generiert ist, während ein wissenschaftlicher Artikel oder ein akademischer Text normalerweise eine präzisere und detailliertere Struktur aufweist. Das ist keine unfehlbare Methode, aber ich denke, dass die Beachtung eines einheitlichen Wortschatzes, der Klarheit und der erkennbaren Komplexität des Inhalts dabei helfen kann, festzustellen, ob er von KI erstellt wurde. Es ist auch sinnvoll, den Inhalt mit dem eigenen Wissen über das Thema abzugleichen. Wenn etwas zu „perfekt“ oder übermäßig neutral erscheint, könnte das ein Hinweis auf eine automatisierte Erstellung sein.Xavi, Spanien, Psychologiestudent, KI-Enthusiast
Leer en español (ES)
Es difícil, especialmente porque no conozco todos los estilos que ChatGPT puede generar. Sin embargo, sí creo que existen ciertas pistas que pueden ayudar. Por ejemplo, el vocabulario que utiliza la IA tiende a ser bastante estable y uniforme. Además, puedes fijarte en la «profundidad» del contenido: un trabajo de clase o un texto más superficial podría despertar sospechas de haber sido generado por IA, mientras que un artículo científico o un texto académico suele tener una estructura más precisa y detallada. No es un método infalible, pero creo que prestar atención a la uniformidad del vocabulario, la claridad y la aparente complejidad del contenido puede ayudar a identificar si ha sido creado por IA. También es útil compararlo con tus propios conocimientos sobre el tema; si algo parece demasiado «perfecto» o excesivamente neutral, podría ser una señal de generación automática.Xavi, España, estudiante de Psicología, entusiasta de la IA
Прочетете на български (BG)
Трудно е, особено защото не познавам всички стилове, които ChatGPT може да генерира. Въпреки това смятам, че има определени признаци, които могат да помогнат. Например речникът, използван от ИИ, обикновено е доста постоянен и еднообразен. Освен това може да се обърне внимание на „дълбочината“ на съдържанието: училищна задача или по-повърхностен текст могат да породят съмнения, че са генерирани от ИИ, докато научна статия или академичен текст обикновено имат по-прецизна и подробна структура. Това не е безпогрешен метод, но смятам, че обръщането на внимание на последователността в речника, яснотата и привидната сложност на съдържанието може да помогне да се установи дали то е създадено от ИИ. Полезно е също така да го сравните със собствените си познания по темата. Ако нещо изглежда прекалено „перфектно“ или твърде неутрално, това може да е признак за автоматично генериране.Xavi, Испания, студент по психология, ентусиаст в областта на ИИ
Bettina Hiller — Text, images and deepfakes: What clues should you look for?
The distinction between human and AI-created content is becoming increasingly difficult as the quality of AI continues to improve. Nevertheless, there are some characteristics and strategies that can help to identify differences – both in texts and in images and videos.
With texts, you can pay attention to style and structure. AI texts are often very clear, grammatically correct and logically structured, but sometimes seem a bit “smooth” or neutral. Human texts, on the other hand, more often contain personal opinions, small mistakes, emotional nuances or individual expressions. Another indication can be whether the text provides a lot of information very quickly, but contains little personal experience or concrete examples.
There are also typical hints for pictures. While AI-generated images can be realistic, they sometimes have small inconsistencies – such as unusual finger positions, distorted details, or inconsistent lighting conditions. Texts in images are also often incorrect or difficult to read. Human-made photos, on the other hand, tend to look more natural and consistent.
In the case of videos, the distinction becomes even more complex. Here you should pay particular attention to facial expressions, lip movements and transitions. AI-generated videos (e.g., deepfakes) may have slight delays between sound and lip movement, or may show unnatural facial features. Atypical movements or too perfect smoothness can also be an indication. In addition to these observations, there are also technical tools such as AI detectors or metadata analyses. Nevertheless, these are not always reliable. That’s why it’s important to think critically, check sources and question content. Overall, a combination of attention, experience, and skepticism is the best way to distinguish between human and AI-created content.
Dr. Bettina Hiller, project manager, Austria, AI Enthusiast
Læs på dansk (DK)
Det bliver stadig vanskeligere at skelne mellem menneskeskabt og AI-genereret indhold, efterhånden som kvaliteten af AI fortsætter med at forbedre sig. Ikke desto mindre findes der nogle karakteristika og strategier, som kan hjælpe med at identificere forskellene – både i tekster, billeder og videoer. Når det gælder tekster, kan man være opmærksom på stil og struktur. AI-genererede tekster er ofte meget tydelige, grammatisk korrekte og logisk opbyggede, men kan nogle gange virke lidt for »glatte« eller neutrale. Menneskeskabte tekster indeholder derimod oftere personlige holdninger, små fejl, følelsesmæssige nuancer eller individuelle formuleringer. Et andet tegn kan være, at teksten hurtigt præsenterer en stor mængde information, men indeholder få personlige erfaringer eller konkrete eksempler. Der findes også typiske tegn i billeder. Selvom AI-genererede billeder kan være realistiske, indeholder de nogle gange små uoverensstemmelser – såsom usædvanlige fingerstillinger, forvrængede detaljer eller inkonsekvent belysning. Tekst i billeder er også ofte forkert eller vanskelig at læse. Fotografier taget af mennesker har derimod en tendens til at se mere naturlige og sammenhængende ud. Når det gælder videoer, bliver det endnu mere kompliceret at skelne mellem de to typer indhold. Her bør man især være opmærksom på ansigtsudtryk, læbebevægelser og overgange. AI-genererede videoer (f.eks. deepfakes) kan have små forsinkelser mellem lyd og læbebevægelser eller vise unaturlige ansigtstræk. Usædvanlige bevægelser eller en alt for perfekt glidende bevægelse kan også være tegn. Ud over disse observationer findes der også tekniske værktøjer såsom AI-detektorer eller analyser af metadata. Disse er dog ikke altid pålidelige. Derfor er det vigtigt at tænke kritisk, kontrollere kilder og stille spørgsmål ved indholdet. Samlet set er en kombination af opmærksomhed, erfaring og skepsis den bedste måde at skelne mellem menneskeskabt og AI-genereret indhold på.Dr. Bettina Hiller, projektleder, Østrig, AI-entusiast
Auf Deutsch lesen (DE)
Die Unterscheidung zwischen menschlich erstellten und KI-generierten Inhalten wird immer schwieriger, da sich die Qualität der KI kontinuierlich verbessert. Dennoch gibt es einige Merkmale und Strategien, die dabei helfen können, Unterschiede zu erkennen – sowohl bei Texten als auch bei Bildern und Videos. Bei Texten kann man auf Stil und Struktur achten. KI-generierte Texte sind oft sehr klar, grammatikalisch korrekt und logisch aufgebaut, wirken aber manchmal etwas zu „glatt“ oder neutral. Menschliche Texte enthalten dagegen häufiger persönliche Meinungen, kleine Fehler, emotionale Nuancen oder individuelle Ausdrucksweisen. Ein weiterer Hinweis kann sein, dass ein Text sehr schnell viele Informationen vermittelt, aber nur wenige persönliche Erfahrungen oder konkrete Beispiele enthält. Auch bei Bildern gibt es typische Hinweise. Obwohl KI-generierte Bilder realistisch wirken können, weisen sie manchmal kleine Unstimmigkeiten auf – beispielsweise ungewöhnliche Fingerstellungen, verzerrte Details oder uneinheitliche Lichtverhältnisse. Texte innerhalb von Bildern sind ebenfalls häufig fehlerhaft oder schwer lesbar. Von Menschen aufgenommene Fotos wirken dagegen meist natürlicher und stimmiger. Bei Videos wird die Unterscheidung noch komplexer. Hier sollte man besonders auf Gesichtsausdrücke, Lippenbewegungen und Übergänge achten. KI-generierte Videos (z. B. Deepfakes) können leichte Verzögerungen zwischen Ton und Lippenbewegungen aufweisen oder unnatürliche Gesichtszüge zeigen. Auch ungewöhnliche Bewegungen oder eine übermäßig perfekte Flüssigkeit der Bewegungsabläufe können Hinweise sein. Neben diesen Beobachtungen gibt es auch technische Hilfsmittel wie KI-Detektoren oder Metadatenanalysen. Allerdings sind diese nicht immer zuverlässig. Deshalb ist es wichtig, kritisch zu denken, Quellen zu überprüfen und Inhalte zu hinterfragen. Insgesamt ist eine Kombination aus Aufmerksamkeit, Erfahrung und Skepsis der beste Weg, um zwischen menschlich erstellten und KI-generierten Inhalten zu unterscheiden.Dr. Bettina Hiller, Projektmanagerin, Österreich, KI-Enthusiastin
Leer en español (ES)
Distinguir entre el contenido creado por personas y el generado por IA es cada vez más difícil a medida que la calidad de la inteligencia artificial continúa mejorando. No obstante, existen algunas características y estrategias que pueden ayudar a identificar las diferencias, tanto en textos como en imágenes y vídeos. En el caso de los textos, puedes prestar atención al estilo y a la estructura. Los textos generados por IA suelen ser muy claros, gramaticalmente correctos y estar organizados de manera lógica, pero a veces parecen demasiado «pulidos» o neutrales. Los textos escritos por personas, en cambio, suelen incluir con mayor frecuencia opiniones personales, pequeños errores, matices emocionales o expresiones individuales. Otro indicio puede ser que el texto proporcione mucha información rápidamente, pero incluya pocas experiencias personales o ejemplos concretos. También existen señales típicas en las imágenes. Aunque las imágenes generadas por IA pueden ser realistas, a veces presentan pequeñas incoherencias, como posiciones extrañas de los dedos, detalles deformados o condiciones de iluminación inconsistentes. Los textos que aparecen dentro de las imágenes también suelen contener errores o resultar difíciles de leer. Las fotografías tomadas por personas, por el contrario, tienden a parecer más naturales y coherentes. En el caso de los vídeos, la distinción se vuelve todavía más compleja. Aquí conviene prestar especial atención a las expresiones faciales, los movimientos de los labios y las transiciones. Los vídeos generados por IA (por ejemplo, los deepfakes) pueden presentar pequeños desfases entre el sonido y el movimiento de los labios, o mostrar rasgos faciales poco naturales. Los movimientos atípicos o una fluidez demasiado perfecta también pueden ser indicios. Además de estas observaciones, existen herramientas técnicas como los detectores de IA o los análisis de metadatos. Sin embargo, no siempre son fiables. Por eso es importante pensar de manera crítica, comprobar las fuentes y cuestionar el contenido. En general, combinar la atención, la experiencia y el escepticismo es la mejor manera de distinguir entre el contenido creado por personas y el generado por IA.Dr. Bettina Hiller, directora de proyectos, Austria, entusiasta de la IA
Прочетете на български (BG)
Разграничаването между съдържание, създадено от хора, и съдържание, генерирано от ИИ, става все по-трудно, тъй като качеството на изкуствения интелект продължава да се подобрява. Въпреки това съществуват някои характеристики и стратегии, които могат да помогнат за откриването на разликите – както при текстовете, така и при изображенията и видеоклиповете. При текстовете можете да обърнете внимание на стила и структурата. Текстовете, генерирани от ИИ, често са много ясни, граматически правилни и логично структурирани, но понякога изглеждат прекалено „изгладени“ или неутрални. Текстовете, написани от хора, от друга страна, по-често съдържат лични мнения, малки грешки, емоционални нюанси или индивидуални изрази. Друг признак може да бъде, че текстът предоставя много информация за кратко време, но съдържа малко личен опит или конкретни примери. Съществуват и типични признаци при изображенията. Макар че изображенията, генерирани от ИИ, могат да изглеждат реалистични, понякога съдържат малки несъответствия – например необичайни позиции на пръстите, изкривени детайли или непоследователно осветление. Текстовете в изображенията също често са неправилни или трудни за разчитане. Снимките, направени от хора, от друга страна, обикновено изглеждат по-естествени и последователни. При видеоклиповете разграничаването става още по-сложно. Тук трябва да обърнете особено внимание на израженията на лицето, движенията на устните и преходите. Видеоклиповете, генерирани от ИИ (например дийпфейковете), могат да имат леки разминавания между звука и движенията на устните или да показват неестествени черти на лицето. Необичайните движения или прекалено съвършената плавност също могат да бъдат признак. Освен тези наблюдения съществуват и технически инструменти, като детектори за ИИ или анализи на метаданни. Въпреки това те невинаги са надеждни. Затова е важно да мислим критично, да проверяваме източниците и да поставяме съдържанието под съмнение. Като цяло комбинацията от внимание, опит и скептицизъм е най-добрият начин да различаваме съдържанието, създадено от хора, от това, генерирано от ИИ.Dr. Bettina Hiller, ръководител на проекти, Австрия, ентусиастка в областта на ИИ
Christian Vogel — Beyond visual clues: Why AI detection needs more than spotting mistakes
The clear distinction between human-created and AI-generated content is becoming increasingly difficult. Characteristics that were once considered reliable indicators—such as typical image artifacts or certain writing patterns—are often already outdated. A well-known example is the use of em dashes in text, which for a time were seen as a sign of AI-generated writing. Today, however, many people intentionally use such stylistic elements themselves. As a result, these kinds of markers are no longer sufficient for identifying AI-generated content.
From a regulatory perspective, initial measures are emerging. The EU AI Act includes transparency and labeling requirements for AI-generated content in certain contexts. Technical solutions such as invisible watermarking systems like SynthID may also help. However, they do not provide complete protection or a foolproof method of detection.
This raises a more fundamental question: does every form of AI-generated content need to be clearly distinguishable from human-created work? In many situations, what matters most is the quality of the outcome. When someone performs a mathematical calculation, for example, few people ask whether it was completed by hand or with a calculator. What matters first and foremost is whether the result is correct.
At the same time, these developments are likely to have significant societal consequences. Creative professions, media production, education, and copyright law will all be affected and, in some cases, may need to be redefined. As with previous waves of industrial and technological change, tasks often do not disappear entirely but instead evolve and shift in nature. The key question, therefore, is not whether this transformation will occur, but how we choose to shape and manage it as a society.
Christian Vogel, teacher, Germany, AI Expert and Enthusiast
Læs på dansk (DK)
Det bliver stadig vanskeligere tydeligt at skelne mellem menneskeskabt og AI-genereret indhold. Kendetegn, der tidligere blev betragtet som pålidelige indikatorer – såsom typiske billedfejl eller bestemte skrivemønstre – er ofte allerede forældede. Et velkendt eksempel er brugen af lange tankestreger i tekster, som i en periode blev opfattet som et tegn på AI-genereret tekst. I dag bruger mange mennesker imidlertid bevidst sådanne stilistiske virkemidler. Derfor er denne type kendetegn ikke længere tilstrækkelige til at identificere AI-genereret indhold. Fra et lovgivningsmæssigt perspektiv begynder de første tiltag at tage form. EU’s AI-forordning indeholder krav om gennemsigtighed og mærkning af AI-genereret indhold i bestemte sammenhænge. Tekniske løsninger såsom usynlige digitale vandmærker, eksempelvis SynthID, kan også være en hjælp. De giver dog hverken fuldstændig beskyttelse eller en ufejlbarlig metode til at opdage AI-genereret indhold. Det rejser et mere grundlæggende spørgsmål: Skal alle former for AI-genereret indhold nødvendigvis kunne skelnes tydeligt fra menneskeskabt arbejde? I mange situationer er det vigtigste kvaliteten af resultatet. Når nogen eksempelvis udfører en matematisk beregning, er der kun få, der spørger, om den blev udført i hånden eller med en lommeregner. Det afgørende er først og fremmest, om resultatet er korrekt. Samtidig vil denne udvikling sandsynligvis få betydelige samfundsmæssige konsekvenser. Kreative erhverv, medieproduktion, uddannelse og ophavsretslovgivning vil alle blive påvirket og kan i nogle tilfælde få behov for at blive omdefineret. Ligesom ved tidligere bølger af industriel og teknologisk forandring forsvinder opgaver ofte ikke fuldstændigt, men udvikler sig og ændrer karakter. Det centrale spørgsmål er derfor ikke, om denne forandring vil finde sted, men hvordan vi som samfund vælger at forme og håndtere den.Christian Vogel, lærer, Tyskland, AI-ekspert og -entusiast
Auf Deutsch lesen (DE)
Die klare Unterscheidung zwischen menschlich erstellten und KI-generierten Inhalten wird zunehmend schwieriger. Merkmale, die früher als zuverlässige Hinweise galten – etwa typische Bildartefakte oder bestimmte Schreibmuster –, sind häufig bereits überholt. Ein bekanntes Beispiel ist die Verwendung von Gedankenstrichen in Texten, die eine Zeit lang als Hinweis auf KI-generierte Texte angesehen wurde. Heute verwenden jedoch viele Menschen solche Stilmittel ganz bewusst selbst. Daher reichen derartige Merkmale nicht mehr aus, um KI-generierte Inhalte zu identifizieren. Aus regulatorischer Sicht zeichnen sich erste Maßnahmen ab. Die KI-Verordnung der EU (AI Act) enthält in bestimmten Zusammenhängen Transparenz- und Kennzeichnungspflichten für KI-generierte Inhalte. Auch technische Lösungen wie unsichtbare digitale Wasserzeichen, beispielsweise SynthID, können dabei helfen. Allerdings bieten sie weder einen vollständigen Schutz noch eine absolut zuverlässige Erkennungsmethode. Das wirft eine grundlegendere Frage auf: Muss jede Form von KI-generiertem Inhalt eindeutig von menschlich erstellten Werken unterscheidbar sein? In vielen Situationen kommt es vor allem auf die Qualität des Ergebnisses an. Wenn jemand beispielsweise eine mathematische Berechnung durchführt, fragen nur wenige Menschen, ob diese von Hand oder mit einem Taschenrechner erledigt wurde. Entscheidend ist in erster Linie, ob das Ergebnis korrekt ist. Gleichzeitig werden diese Entwicklungen voraussichtlich erhebliche gesellschaftliche Auswirkungen haben. Kreative Berufe, Medienproduktion, Bildung und Urheberrecht werden davon betroffen sein und müssen in manchen Fällen möglicherweise neu definiert werden. Wie bei früheren Wellen des industriellen und technologischen Wandels verschwinden Aufgaben häufig nicht vollständig, sondern entwickeln sich weiter und verändern ihren Charakter. Die entscheidende Frage ist daher nicht, ob dieser Wandel stattfinden wird, sondern wie wir ihn als Gesellschaft gestalten und bewältigen wollen.Christian Vogel, Lehrer, Deutschland, KI-Experte und -Enthusiast
Leer en español (ES)
Cada vez resulta más difícil distinguir claramente entre el contenido creado por personas y el generado por IA. Las características que antes se consideraban indicadores fiables, como los errores típicos en las imágenes o determinados patrones de escritura, a menudo ya han quedado obsoletas. Un ejemplo conocido es el uso de rayas largas en los textos, que durante un tiempo se consideró una señal de escritura generada por IA. Sin embargo, hoy en día muchas personas utilizan intencionadamente estos recursos estilísticos. Como consecuencia, este tipo de indicios ya no son suficientes para identificar el contenido generado por IA. Desde el punto de vista normativo, están empezando a surgir las primeras medidas. El Reglamento de Inteligencia Artificial de la Unión Europea (AI Act) incluye requisitos de transparencia y etiquetado para el contenido generado por IA en determinados contextos. También pueden ayudar soluciones técnicas como los sistemas de marcas de agua invisibles, por ejemplo SynthID. Sin embargo, no ofrecen una protección completa ni un método infalible de detección. Esto plantea una cuestión más fundamental: ¿es necesario que todas las formas de contenido generado por IA puedan distinguirse claramente de las obras creadas por personas? En muchas situaciones, lo que más importa es la calidad del resultado. Por ejemplo, cuando alguien realiza un cálculo matemático, pocas personas preguntan si lo ha hecho a mano o con una calculadora. Lo fundamental es, ante todo, que el resultado sea correcto. Al mismo tiempo, es probable que estos avances tengan importantes consecuencias sociales. Las profesiones creativas, la producción de contenidos audiovisuales, la educación y la legislación sobre derechos de autor se verán afectadas y, en algunos casos, quizá deban redefinirse. Como ocurrió con anteriores oleadas de cambios industriales y tecnológicos, las tareas a menudo no desaparecen por completo, sino que evolucionan y cambian de naturaleza. Por lo tanto, la cuestión fundamental no es si esta transformación tendrá lugar, sino cómo decidimos darle forma y gestionarla como sociedad.Christian Vogel, profesor, Alemania, experto y entusiasta de la IA
Прочетете на български (BG)
Ясното разграничаване между съдържание, създадено от хора, и съдържание, генерирано от ИИ, става все по-трудно. Характеристики, които някога са се смятали за надеждни признаци – като типични дефекти в изображенията или определени модели на писане – често вече са остарели. Добре известен пример е използването на дълги тирета в текстовете, които за известно време се приемаха като признак за текст, генериран от ИИ. Днес обаче много хора съзнателно използват такива стилистични средства. В резултат на това подобни признаци вече не са достатъчни за разпознаване на съдържание, генерирано от ИИ. От регулаторна гледна точка започват да се появяват първите мерки. Регламентът на ЕС за изкуствения интелект (AI Act) включва изисквания за прозрачност и обозначаване на съдържание, генерирано от ИИ, в определени ситуации. Технически решения като системите за невидими цифрови водни знаци, например SynthID, също могат да помогнат. Те обаче не осигуряват пълна защита или безпогрешен метод за откриване. Това повдига един по-фундаментален въпрос: необходимо ли е всяка форма на съдържание, генерирано от ИИ, да може ясно да се различава от произведенията, създадени от хора? В много ситуации най-важно е качеството на резултата. Например, когато някой извършва математическо изчисление, малко хора питат дали то е направено на ръка или с калкулатор. Най-важното преди всичко е дали резултатът е правилен. В същото време тези развития вероятно ще имат значителни обществени последици. Творческите професии, медийното производство, образованието и законодателството за авторското право ще бъдат засегнати и в някои случаи може да се наложи да бъдат предефинирани. Както при предишните вълни на индустриални и технологични промени, задачите често не изчезват напълно, а се развиват и променят своя характер. Следователно ключовият въпрос не е дали тази трансформация ще се случи, а как ние като общество ще изберем да я оформим и управляваме.Christian Vogel, учител, Германия, експерт и ентусиаст в областта на ИИ
Rumen Neonia — Can we trust what we see? The need for digital transparency and media literacy
It’s going to get harder and harder. We need technical standards for content provenance—transparency and basic media literacy.
Rumen Neonia, an organizational psychologist, Bulgaria, AI Educator
Læs på dansk (DK)
Det bliver sværere og sværere. Vi har brug for tekniske standarder til at dokumentere indholdets oprindelse – gennemsigtighed og grundlæggende medieforståelse.Rumen Neonia, organisationspsykolog, Bulgarien, AI-underviser
Auf Deutsch lesen (DE)
Es wird immer schwieriger werden. Wir brauchen technische Standards zur Nachverfolgbarkeit der Herkunft von Inhalten – Transparenz und grundlegende Medienkompetenz.Rumen Neonia, Organisationspsychologe, Bulgarien, KI-Pädagoge
Leer en español (ES)
Será cada vez más difícil. Necesitamos estándares técnicos que permitan verificar el origen de los contenidos, transparencia y conocimientos básicos de alfabetización mediática.Rumen Neonia, psicólogo organizacional, Bulgaria, educador en IA
Прочетете на български (BG)
Ще става все по-трудно. Нуждаем се от технически стандарти за проследяване на произхода на съдържанието – прозрачност и основна медийна грамотност.Rumen Neonia, организационен психолог, България, преподавател по ИИ
Stefan Grigorov — Too perfect to be real? Spotting AI avatars and artificial scenes
Ai avatars are to perfect, the atmosphere is to harmonious.
Stefan Grigorov, creator of the platform Epoha online, Bulgaria, AI Educator
Læs på dansk (DK)
AI-avatarer er for perfekte, og atmosfæren er for harmonisk.Stefan Grigorov, skaber af platformen Epoha online, Bulgarien, AI-underviser
Auf Deutsch lesen (DE)
KI-Avatare sind zu perfekt, und die Atmosphäre ist zu harmonisch.Stefan Grigorov, Gründer der Plattform Epoha online, Bulgarien, KI-Pädagoge
Leer en español (ES)
Los avatares generados por IA son demasiado perfectos y el ambiente es demasiado armonioso.Stefan Grigorov, creador de la plataforma Epoha online, Bulgaria, educador en IA
Прочетете на български (BG)
Аватарите, генерирани от ИИ, са прекалено съвършени, а атмосферата е прекалено хармонична.Stefan Grigorov, създател на платформата Epoha online, България, преподавател по ИИ
You’ve taken your first steps with AI. Ready to explore what the future might bring?
Continue your journey by discovering how AI could transform our work, education and everyday lives—and the opportunities and challenges that may lie ahead.
That is a challenge that is going to increase its difficulty progressively in the incoming years. As the generative AI will be trained more and better and the models will improve, the small signs of AI generated content are going to become almost impossible to distinguish. So I would recommend, as the first measure to be sure of the content’s origins, to not trust any content we can find online without a proper human signature. And even if that firm is there, maybe it would be healthy to still doubt its source. We are, sadly, entering a time in human history where a text or photography is not a proof of facts anymore. For the most empirical approach I would point out some signs that have helped me make a distinction between AI generated content and the human one. Text-wise, language models usually go for short enunciative sentences with not much rhetorical figures and self-explanatory points of view. Either if they are supporting or criticizing any point, they won’t mix different povs. Talking about images or videos, usually they have some problems filling pixels in complex structures, so if the observer is able to magnify the image it would be easy to spot some inconsistency. As I say, most of these issues are being fixed by many models right now, so it’s a problem every part of society will need to attend to at some point with no clear outcome.
This distinction has become harder as AI tools have improved during the past few years, and it will likely only become even harder in the future. As of right now text is almost impossible to detect for being AI made just by looking at it. Image and video content have also improved significantly in recent years, but video still looks to be the most challenging to generate with AI. Especially video content including humans or animals can be more easily detected due to all the mistakes that can occur with these categories. However, to answer the question “How can we distinguish”, the true answer probably is that we can’t. Especially when video generation improves further. We can hope that AI generated content will be marked as AI generated, but it will definitely be an easy way to spread misinformation. Therefore, the only thing we can really do is to try seeking trustworthy sources and not believe everything we see on a random video on the internet.
I believe that, with experience, you start to notice certain patterns. Sometimes, content generated by artificial intelligence can have a style that is too structured, very uniform, or even slightly impersonal, and these characteristics can serve as clues that it was AI-generated. Recognizing these subtle hints can help users develop a better sense of when AI might be involved. However, beyond these indicators, it is not always easy to distinguish AI-generated content, to be honest. For this reason, it may be most useful to rely on more sophisticated or specialized tools designed specifically to detect AI-generated material. Even with these tools, though, the task is becoming increasingly complex, and I believe that over time it will be even more difficult to differentiate between human-produced and AI-generated content. As AI technology continues to advance, the lines are likely to blur further, making critical assessment and verification even more essential.
It's difficult, specially because I'm not familiar with all the styles that ChatGPT can generate. However, I do think there are certain clues that can help. For example, the vocabulary used by AI tends to be quite stable and uniform. In addition, you can look at the "depth" of the content: a class assignment or a more superficial text might raise suspicious of being AI-generated, whereas a scientific article or an academic text usually has a more precise and detailed structure. It's not a fool-proof method, but I think paying attention to consistency in vocabulary, clarity, and the apparent complexity of the content can help identify whether it has been created by AI. It's also useful to compare it with your own knowledge of the topic; if something seems too "perfect" or overly neutral, it could be a sign of automated generation.
The distinction between human and AI-created content is becoming increasingly difficult as the quality of AI continues to improve. Nevertheless, there are some characteristics and strategies that can help to identify differences – both in texts and in images and videos. With texts, you can pay attention to style and structure. AI texts are often very clear, grammatically correct and logically structured, but sometimes seem a bit "smooth" or neutral. Human texts, on the other hand, more often contain personal opinions, small mistakes, emotional nuances or individual expressions. Another indication can be whether the text provides a lot of information very quickly, but contains little personal experience or concrete examples. There are also typical hints for pictures. While AI-generated images can be realistic, they sometimes have small inconsistencies – such as unusual finger positions, distorted details, or inconsistent lighting conditions. Texts in images are also often incorrect or difficult to read. Human-made photos, on the other hand, tend to look more natural and consistent. In the case of videos, the distinction becomes even more complex. Here you should pay particular attention to facial expressions, lip movements and transitions. AI-generated videos (e.g., deepfakes) may have slight delays between sound and lip movement, or may show unnatural facial features. Atypical movements or too perfect smoothness can also be an indication. In addition to these observations, there are also technical tools such as AI detectors or metadata analyses. Nevertheless, these are not always reliable. That's why it's important to think critically, check sources and question content. Overall, a combination of attention, experience, and skepticism is the best way to distinguish between human and AI-created content.
The clear distinction between human-created and AI-generated content is becoming increasingly difficult. Characteristics that were once considered reliable indicators—such as typical image artifacts or certain writing patterns—are often already outdated. A well-known example is the use of em dashes in text, which for a time were seen as a sign of AI-generated writing. Today, however, many people intentionally use such stylistic elements themselves. As a result, these kinds of markers are no longer sufficient for identifying AI-generated content. From a regulatory perspective, initial measures are emerging. The EU AI Act includes transparency and labeling requirements for AI-generated content in certain contexts. Technical solutions such as invisible watermarking systems like SynthID may also help. However, they do not provide complete protection or a foolproof method of detection. This raises a more fundamental question: does every form of AI-generated content need to be clearly distinguishable from human-created work? In many situations, what matters most is the quality of the outcome. When someone performs a mathematical calculation, for example, few people ask whether it was completed by hand or with a calculator. What matters first and foremost is whether the result is correct. At the same time, these developments are likely to have significant societal consequences. Creative professions, media production, education, and copyright law will all be affected and, in some cases, may need to be redefined. As with previous waves of industrial and technological change, tasks often do not disappear entirely but instead evolve and shift in nature. The key question, therefore, is not whether this transformation will occur, but how we choose to shape and manage it as a society.
It’s going to get harder and harder. We need technical standards for content provenance—transparency and basic media literacy.
Ai avatars are to perfect, the atmosphere is to harmonious
