Can teachers spot AI writing? What Penn researchers learned - and what writers should do
AI is everywhere in classrooms and creative briefs. That brings a messy question to the surface: can anyone reliably tell whether a piece was written by a human or a model?
Recent work led by University of Pennsylvania's Chris Callison-Burch and PhD student Liam Dugan offers a blunt answer: detection is shaky, and high-stakes accusations are risky.
What the research actually shows
Commercial AI detectors look confident on marketing pages. In practice, they're inconsistent. Using a massive dataset (RAID) with more than 10 million documents, the Penn team found detectors do alright on copy-pasted model output - but their accuracy drops once text is edited, paraphrased, reordered, or blended with human writing.
Worse, false positives happen. Rates of 5-6% may sound small until you scale them to a large class or a newsroom. Tighten the threshold to avoid false accusations, and the tools miss even more AI-written text.
There's also history here: even OpenAI stepped back from an AI-text classifier due to low accuracy, noting it shouldn't be used for decisions that affect people's outcomes. Read their note.
Humans aren't great detectors either
Studies show people are barely better than chance at separating AI from human writing. Training helps, but only with strong incentives and feedback loops. As models improve, the line blurs further, and "gut feel" becomes even less reliable.
What this means if you're a writer
Whether you use AI or not, the risk isn't abstract. Editors, clients, and platforms are experimenting with detectors. False flags can cost you money and reputation. Here's how to protect your work and stay credible.
Protect your credibility (and income)
- Document your process. Keep drafts, timestamps, version history, and research notes. Screenshots of your workflow beat a detector score every time.
- Set expectations in contracts. Define acceptable AI use (idea generation, outlines, editing) vs. what's off-limits. Specify that detector outputs aren't grounds for nonpayment without corroborating evidence.
- Disclose with care. If you use AI, state where and how. Add a brief note in deliverables when requested. Transparency builds trust and diffuses suspicion.
- Keep a human fingerprint. Personal anecdotes, original interviews, proprietary data, and unique examples signal authorship better than style flourishes.
- Own the edits. If AI helps draft, run a ruthless human pass: restructure, fact-check, add voice, and tailor for the brief. Make it unmistakably yours.
- Push back on "detector verdicts." If flagged, ask for specifics, provide your draft history, and request a human editorial review instead of automated scores.
Ethics that actually hold up
- Don't pass off raw model output as final. It undercuts your value and invites quality issues.
- Verify facts and sources. Models hallucinate. Your name is on the line.
- Respect originality. Blend research with lived experience, interviews, and client-specific insights. That's where your advantage sits.
For editors and clients evaluating submissions
- Use detectors as signals, not verdicts. If something feels off, request drafts and notes, then assess the piece on quality and originality.
- Reward process, not just output. Ask for outlines, sources, and a short "how I approached this" note.
- Write clear AI policies. Define acceptable support (brainstorming, grammar) vs. banned uses (undisclosed, wholesale generation).
Key takeaways from Penn's findings
- AI detectors aren't reliable enough for high-stakes decisions, especially with edited text.
- False positives are real and consequential.
- Humans can improve at spotting AI with training, but accuracy remains limited as models improve.
- Accusations based on "gut feel" or a single tool can lead to unfair outcomes.
Practical next steps for writers
- Create a versioned workflow (outline → draft → revisions) and store it.
- Write a one-paragraph AI disclosure you can reuse on request.
- Audit your unique voice: add stories, frameworks you coined, and owned data.
- Update contracts to address AI use and dispute resolution beyond detector scores.
If you want structured training on ethical, high-quality AI-assisted writing, explore these resources:
Bottom line: treat detectors as rough signals, not judges. Protect your process, be transparent, and keep delivering work only a human with context, taste, and judgment can ship.
Your membership also unlocks:
