Unique Research · Thinking · December 31, 2024
Over the past year you may have read or watched work that no human wrote. AI is spreading across text, image, and video. Since ChatGPT launched in November 2022, use has surged. Europol’s Innovation Lab has estimated that by 2026 as much as 90% of online content could be AI-synthesized. Facing that trajectory, knowing what is synthetic — and what it means — matters more than ever.
Spotting AI text
Paul Shovlin, assistant professor of English and AI specialist at Ohio University, says detecting AI text is hard, especially across genres. Teachers may flag work that lacks a student’s usual specificity and diction — harder when writing isn’t personal. “Traits used for gut judgment aren’t stable across genres. A science report won’t have a quirky personal voice.”
False positives happen too: neurodiverse writers have been flagged as likely AI when they used none.
LLMs sometimes leave “tells.” Chad Mourning, Ohio University CS assistant professor, notes LLMs predict the next best word — which can overuse tokens overrepresented in training but rare in speech. “‘Inquire’ shows up a lot in student papers — they wouldn’t say it out loud. That raises suspicion.”
Early LLMs were verbose and didn’t know when to stop; newer ones copy organic text better. Advanced prompting and bot programming can make AI writing look more “organic” than generic ChatGPT defaults.
Spotting AI images
AI often struggles with human features — faces and fingers. Counting fingers or spotting warped faces is a quick check. Odd proportions are red flags. Apps — and AI itself — can help detect AI images.
“In theory any AI image can be detected by AI, but far more effort goes into generation than detection,” Mourning says. “That’s basically GANs: train a generator, train a discriminator on fakes, exclude what the discriminator catches to train a better generator — which then trains a better detector.”
How data and the internet shape outputs
LLMs are shaped by training data. “Most generative algorithms are weighted combinations of training examples — one part this, one part that. If every butterfly photo is symmetric, generated butterflies will be too.”
Biased or bad training yields biased or bad output. Deliberate misinformation in user content can poison models: AI search once suggested chewing rocks for health — based on a humorous Reddit reply. Not always dangerous; sometimes subtler.
Shovlin suggests grounding tools in supplied text only: “Prompt ChatGPT and peers to stick to texts you give them. If the bot doesn’t roam the wider web, you can reasonably trust replies come from your loaded sources.”
Is AI use ethical?
It depends. Mourning and Shovlin say generative AI isn’t inherently unethical; deception and privacy are greyer. Shovlin urges rhetorical awareness: “If my audience knew this text was AI-generated, what would they think? What does my organization expect on privacy, copyright, and AI vs human text?”
How is AI regulated?
Mourning sees the biggest ethics issues as deception and unauthorized training data. Disclosure helps on deception; data is harder. Some LLMs trained on YouTube transcripts without clear creator consent. Forcing full dataset disclosure may reveal trade secrets; listing sources is a compromise so artists and platforms can check rights and ToS.
Shovlin is more pessimistic about meaningful regulation: companies are powerful, tech is complex, politicians often seem thin on tech (see social-media hearings), and “regulation hurts innovation” is a strong lobby given AI’s perceived potential.
Will AI replace writers and creatives?
Shovlin notes AI already displaces some writers — e.g., ESPN controversy over AI stories for niche sports that humans might cover. “Jobs change; we should carefully weigh labor impact and use our voice when needed.”
Creatives are being displaced as AI strengthens — but Mourning says some can use the tools: “There will always be room for creatives, fewer of them. Existing writers may become top early ‘prompt engineers.’ Transformation, not extinction.”
— END —