Student writing

Thesis Ack Scored 1.0 on Two AI Detectors. The Fix Isn't Fancy Words

Your generic gratitude paragraph is a statistical fingerprint of a language model. One hyper-specific detail and a broken sentence rhythm is all it takes to look like a person.

We ran a controlled experiment on thesis acknowledgements. Two independent AI-text detectors, one generic baseline, seven targeted variations, every variable isolated. The results should make every grad student who cranked out an acknowledgement at 2 a.m. before hitting 'submit' genuinely nervous. Your baseline paragraph, the one with 'I would like to express my deepest gratitude for the invaluable guidance,' scores 1.0 and 0.9999. Full AI. Maximum flag. And here is the part that should keep you up: swapping 'grateful' for 'eternally appreciative' and stacking three more transition phrases barely moves the needle. You stay at 0.9945 and 0.9999. The detector does not care about your thesaurus.

Here is the full table, measured, not guessed. We used two independent prob_ai models; lower means more human. Baseline: 1.0 / 0.9999. Transition-word bloat: 0.9945 / 0.9999. Thesaurus fancy vocabulary: 0.9976 / 0.9999. Burstiness, mixed sentence length: 0.0239 / 0.1732. First-person with one concrete detail: 0.0638 / 0.0369. Deliberate imperfection, fragments and hedges: 0.9831 / 0.9811. Deliberate rhythm, short-long beats: 0.0615 / 0.0401. Method: same base text, one variable changed per trial, both detectors scored independently. Two of the results go against every writing-tutorial article you have read, and the other two confirm what you already suspect about your own draft.

First, the most expensive trap. You read ten articles saying 'use richer vocabulary and smoother transitions.' You do it. Your acknowledgement now has 'furthermore,' 'moreover,' 'it is worth noting that,' and 'I remain profoundly indebted to.' Congratulations. Score: 0.9945 / 0.9999. You added eleven words and changed your detection probability by essentially zero. The detector is not reading your word list. It is reading the statistical fingerprint of your sentence lengths, your information density, your predictability. Fancy adjectives are noise to that fingerprint. The output stays flat, uniform, high-entropy, and that is exactly the shape a language model produces when prompted for formal gratitude.

Now the part that actually works. Genuinely mix a three-word sentence with a forty-word run-on, and detector A drops to 0.0239. Near zero. Add one specific first-person detail, not 'my family supported me' but 'my mom called every Sunday for three years and I answered on the first ring exactly twice,' and detector B drops to 0.0369. Deliberate short-long-short rhythm keeps both under 0.07. The pattern: the detector flags statistical uniformity. One concrete, slightly awkward, hyper-specific detail breaks the pattern. It reads like a person remembering a Tuesday, not a model interpolating a gratitude template. If you need a structural quick-fix before you sit down and rewrite from memory, easydue's strong mode takes the raw 1.0 baseline to 0.2494 on detector A. Useful. But it cannot invent your mom's missed Sunday calls for you. That has to come from your actual three years of data.

Here is the counterintuitive result that should rewire your revision strategy. We tested deliberate imperfection: fragments, hedges, a slightly unpolished casual feel. 'Honestly, I could name forty people. I will not. You know who you are.' Sounds human, right? Score: 0.9831 / 0.9811. It made the text more AI-like, not less. Why? Because when a language model is prompted to sound casual and imperfect, it produces a very specific, repeatable pattern of fragments and hedges. You are mimicking the model's imitation of humanity. The fix is not to add performed roughness. It is to write the actual sentence that is in your head, even if it is grammatically uneven, and let the imperfection be accidental rather than staged.

So what does a professional, non-generic acknowledgement actually look like? Three rewrites. ❌ 'I would like to express my deepest gratitude to my supervisor for their invaluable guidance and support throughout this research journey.' ✅ 'Dr. Chen spent a Tuesday at eleven p.m. in the lab because my data pipeline broke for the third time. I will never send him a just-checking-in email on a Friday evening.' ❌ 'I am thankful to my family for their unconditional support.' ✅ 'My mom called every Sunday for three years. I answered on the first ring exactly twice. She does not know I cried in the data room in March. She does not need to.' ❌ 'This research would not have been possible without the contributions of many individuals.' ✅ 'The data-room door on the third floor has been broken since September. I held it for roughly forty grad students. Every one of them became a co-conspirator in my deadline.'

A thesis acknowledgement is the one page in your entire dissertation where you are allowed to be a person. Not a researcher. Not a writer. A person who was exhausted, specific, and a little bit ungraceful. Use that page. If you have already generated your acknowledgement in four minutes and it reads like a LinkedIn post crossed with a wedding toast, easydue will pull you from 1.0 to roughly 0.25. Better. Still not you. The real fix is sitting in your memory: the specific Tuesday, the broken door, the phone call you missed. Write that down. It will score lower on the detector and mean more to the committee than any 'invaluable guidance' paragraph ever will.