AI Detectors Are Getting It Wrong: False Positives, Academic Risk and How to Respond

AI Detector Flags Abraham Lincoln’s Gettysburg Address as AI-Generated: What’s Going On?

A Reddit post claims an AI detector flagged Abraham Lincoln’s Gettysburg Address as AI-generated, and a professor’s 45-year-old paper scored 77% “AI”. It’s not the first time classic or well-edited prose has been mislabelled by automated detectors.

“Colleges are using this to end peoples careers and innocent people get punished.”

Whether you work in UK higher education, publish professionally, or simply write online, the lesson is the same: current AI text detectors are not reliable enough to make high-stakes decisions about authorship.

Source: Reddit thread shared by /u/velorae.

Why AI text detectors produce false positives

Most detectors try to spot “AI-like” statistical patterns such as low perplexity (predictable next words) and consistent sentence structure and phrasing (“burstiness”). That’s a proxy, not proof of authorship. Several factors trip them up:

• Polished, formulaic, or encyclopaedic writing – including historical speeches, policy documents, and edited academic prose – can look “AI-like”.
• Short passages and abstracts lack enough signal for accurate classification, increasing error rates.
• Non-native English or simplified writing may appear “too regular”, creating bias against certain groups.
• Modern models emulate human variability well, while humans can write with machine-like regularity. Boundaries blur.

Even vendors caution against over-reliance. OpenAI discontinued its own AI text classifier, citing low accuracy. Turnitin’s guidance states that AI-writing indicators should not be the sole basis for academic misconduct decisions.

• OpenAI: AI Text Classifier (retired) and follow-up notice on discontinuation.
• Turnitin: AI writing detection – vendor cautions and policy notes.

Academic and professional risk in the UK

In UK universities and workplaces, a false positive can trigger formal investigations, damaged reputations, and stress. That risk is amplified if institutions treat a detector score as definitive truth.

Automated decisions and UK data protection

Under UK GDPR, individuals have protections against decisions based solely on automated processing where those decisions have legal or similarly significant effects. An AI detector flag should therefore be treated as a lead, not a verdict, and must be accompanied by meaningful human review and opportunity for explanation. See the ICO’s guidance on AI and data protection.

UK higher education processes

Universities must follow fair procedures, disclose evidence, and allow students or staff to respond. Sector bodies such as Jisc urge proportionate, transparent assessment practices when using generative AI. See Jisc’s advice on generative AI in education.

If your work is wrongly flagged as AI-generated: a practical response

If you’re a student, researcher, or professional in the UK and you receive an “AI-generated” allegation, here’s a clear, defensible approach.

• Ask for the policy: Request the institution’s written policy on AI usage and detection. Ask how the detector works, known error rates, and validation evidence.
• Insist on human review: Cite vendor cautions and UK GDPR principles on avoiding solely automated decisions. Ask for an academic or subject matter review of the work itself.
• Provide a provenance trail: Share drafts, version history (Google Docs, Word tracked changes, Overleaf, Git commits), notes, outlines, and reading lists. Screenshots and timestamps help.
• Explain your process: Provide a short statement describing how you researched, planned, and wrote the piece. Reference sources and specific choices you made.
• Offer comparative samples: Provide previous writing samples to show consistent voice and style, while noting that stylometry is indicative, not definitive.
• Address privacy: Avoid uploading your full work to third-party detectors without understanding data retention. Ask your institution about approved tools and data handling.
• Seek support: Students can consult their students’ union or advisor; staff can speak to their trade union or HR. Consider the Office of the Independent Adjudicator (OIA) for eligible student complaints after internal processes conclude.

If you do use AI as part of your workflow, keep a transparent log of prompts, outputs, and edits. A simple spreadsheet or document can be enough. For example, you can automate logs with Google Sheets; here’s a practical guide to connecting ChatGPT and Google Sheets to streamline record-keeping.

Guidance for UK institutions and employers

AI detectors can be useful as triage tools, but policy and practice must reflect their limits.

• Never act on a detector score alone. Require corroboration such as process evidence, source use, and a viva or reflective commentary.
• Publish a clear AI policy: What is permitted, how disclosure works, and how evidence is evaluated.
• Complete a DPIA (data protection impact assessment) for any detector use and document lawful basis, fairness, and human-in-the-loop safeguards.
• Mitigate bias risk: Provide routes for non-native English writers and neurodivergent authors to evidence process without penalty.
• Design assessments for authenticity: In-class tasks, oral defences, iterative drafts, personal datasets, and reflective components reduce dependence on unreliable detection.
• Train staff: Ensure investigators understand detector limitations, academic integrity standards, and how to conduct fair reviews.

Better alternatives to “gotcha” detection

There is no foolproof “AI lie detector” for text. More robust approaches focus on process and learning outcomes:

• Provenance by design: Require drafts, research notes, and revision logs as part of submission.
• Targeted vivas: Short, supportive discussions to confirm understanding, not to entrap.
• Transparent AI use: Allow declared, limited use of tools for brainstorming or editing, with critical reflection on contributions and limitations.
• Assessment design: Use tasks that require local data, practical application, or personal synthesis that off-the-shelf models cannot replicate easily.

Bottom line: treat detector outputs as leads, not verdicts

The Gettysburg Address being flagged as “AI” is a neat headline for a real problem: statistical detectors confuse style with authorship. Vendors themselves say not to use scores in isolation. Under UK data protection and good academic practice, people deserve a fair process and meaningful human review.

For individuals: document your writing process and push back, politely but firmly, on uncorroborated accusations. For institutions: build policies, assessment, and training that improve integrity without outsourcing judgement to fragile classifiers.