Accused of Using AI at Work? What to Do Next

A client or manager who says your delivered work reads as machine-written has usually decided something already. The useful question is not whether you win the argument today, but what the decision rests on tomorrow. Four moves come first, and none of them touches the text.

Before those, one document worth having open. The company behind the detector in one of the most widely reported cases in this space has said publicly that its tool should not be used to punish people. Originality.AI’s chief executive, Jonathan Gillham, told Gizmodo in 2024 that his company advises against its tool being used in academia and, in his words, “strongly recommend against being used for disciplinary action.” Read the scope honestly, because he is describing academic discipline rather than employment, and quote it as that. A vendor warning against his own product as a basis for punishing someone is still an unusual thing to be able to hand a manager, and attaching it costs you nothing.

Get the accusation in writing before you answer it. In an academic case there is normally a document to argue with: a report, a score attached to a submission, a file somebody opened. At work the accusation frequently exists only as a sentence said in a meeting or on a call, and a spoken sentence can be softened, sharpened or forgotten later by the person who said it. So if it arrived verbally, put it back in writing yourself: a short message summarising what you were told, by whom, and when. Neutral wording, no defence attached. If the relationship ends over this, your summary may be the only reason anyone can reconstruct what was actually alleged.

Leave the flagged file exactly as it is. Do not open it to smooth a sentence, do not run it through a rewriting tool, and do not send a tidied version to make the number fall. If you wrote the piece, the file as delivered is the strongest object you own, and every edit made after the complaint is a change you will have to account for. Note who is telling you that: I build and sell AI-writing software, so this is the instruction my industry has the least commercial reason to give. Every source below is named and dated, so you can open each one and read it for yourself.

Ask for the tool's name, the number it returned, and the threshold that number is supposed to clear. Ask for the specific passages too. A percentage cannot be answered, because there is nothing in it to answer. Sentences can. The third question often does the most work, because frequently there is no written threshold anywhere, only a reading someone treated as a finding. Ask which tool for a second reason as well: the phrase covers a lot of ground, and some accusers count a grammar checker or a translation aid alongside a chatbot, so the answer tells you what you are actually being accused of using. Where no policy sits behind the number, that absence is the thing to name, calmly and in writing.

Ask that a person, not a score, make the decision. Frame it as a request: that whoever decides weighs your record on the account, the brief, and how the work came together, with the tool’s output as one input rather than the verdict. Someone can grant that without conceding the number was wrong, which makes it the easiest thing on this list to say yes to.

Now the part this space rarely says out loud. School has a ladder: instructor, then a department or integrity office, then a hearing with a written record. Nothing equivalent turned up on the work side in anything read for this page. None of the four freelance platforms quoted below publishes an appeal route for a writer who says the flag is wrong, and no comparable employment process appeared in the legal sources either. So you are not appealing a finding. You are trying to influence a decision one person is entitled to make quickly, which is why the record and the request for human judgement matter more here than the rebuttal does.

If the reply is silence, read it as information rather than insult. Send one follow-up restating the request with a date on it, then spend the remaining energy on the last two sections of this page: the record that already exists, and the published research on how often other detectors are wrong about real people.

Your position turns on a distinction almost nothing written about AI accusations bothers to make: whether you are an employee of the organisation that flagged you, or a contractor delivering work to it. Different bodies of rules, different first moves.

An employee on a W-2 has internal documents to work with. There is a handbook, in many companies some form of AI or acceptable-use policy, a manager, and an HR function keeping records whether or not that helps you. US employment law sets the outer boundary of what can be done to you, and the next section covers what that boundary actually is. The immediate task is to establish whether a written AI policy exists at all, what it says about tools and disclosure, and whether the person who flagged you followed it.

A 1099 contractor is largely outside that frame. At-will termination doctrine does not reach independent contractors, whose relationship is governed by the contract and by whatever statutes apply to contract work instead. I am attributing that to Stop Unpaid Wages, a California employment-law firm publishing it as part of its own marketing, which makes it a self-interested source for a claim that is nevertheless standard US common-law structure. The consequence is blunt: your protection is the document you signed plus whatever the platform’s terms add. If neither mentions AI, neither addresses this dispute.

Most say nothing. Freelancers Union’s own downloadable contract template, the concrete artefact its contract resources point at, carries no AI-disclosure, AI-warranty or AI-conduct language at all; a text extraction returns zero matches for artificial intelligence, generative or ChatGPT. The National Writers Union does have an official generative-AI platform, adopted in October 2023, but read it before relying on it. It protects writers from having their work used to train models, states that “Creators should not be required by employers or clients to use generative AI in our work,” and requires employers to disclose when materials handed to a writer are themselves AI-generated. Every provision runs opposite to your problem. None of it helps a writer who needs to show they did not use AI.

The closest real model clause has the same inversion built in. The Authors Guild publishes model AI clauses for book contracts in which the disclosure obligation flows from author to publisher: the author “shall disclose to Publisher if any AI-generated text is included in the submitted manuscript,” subject to a de minimis cap. Put plainly, the only published model AI clause this research could find is a rule for an author declaring AI use to a publisher, which is the reverse of what a freelance writer accused by a client needs, since the freelancer has to demonstrate that no AI was used at all. If you want something in your next contract, the drafting problem is a warranty that the work is your own plus an agreed procedure for what happens if either side wants it checked, and that is a conversation for a lawyer rather than a paragraph pasted from a page like this.

One gap I will not paper over. If the client is a direct contract rather than a platform and the accusation arrives attached to an unpaid invoice, nothing documenting how that specifically plays out turned up in this research. The remedies for an unpaid invoice are the ordinary ones and are not particular to this dispute. Anyone describing an AI-specific procedure for getting paid after a detector flag is describing something that, as far as this sweep could establish, does not exist in published form.

Everything in this section is United States law and applies nowhere else. Most other jurisdictions run some form of unfair-dismissal protection with no at-will equivalent, so if you are reading from outside the US, this analysis is not yours and a local employment lawyer is the shortest path.

At-will employment is the US default, and the honest answer to the heading is yes: in most at-will arrangements an employer can end the relationship over an accusation like this. Cornell Law School’s Wex entry on the doctrine names three judicial exceptions to it, the public-policy exception, the implied-contract exception, and the implied covenant of good faith and fair dealing, which Cornell notes is recognised in some states, California among them. Cornell declines to say how many states recognise each, stating only that “These wrongful termination exceptions will typically vary by state.” Sources printing state counts do circulate, and the ones checked here disagreed with each other, so no count appears on this page. Yours is a question for a lawyer licensed where you work.

Now the part that has to be labelled. No source found for this page connects any of those three exceptions to a firing over an AI-authorship accusation, and reasoning from the doctrine to your facts is my inference rather than any court’s holding. An implied-contract argument only becomes interesting if your employer published a written procedure and then skipped it, and whether that argument works anywhere is exactly what has not been tested.

It has not been tested because, as far as the public record shows, nobody has brought the case. A dedicated tracker of AI-detection lawsuits, plus targeted searching around it, produced no employment case at all: no worker suing over being falsely accused of submitting AI-generated work. Every documented matter is a student against an institution, including Newby v. Adelphi, Doe v. Yale School of Management, Yang v. University of Minnesota, Harris v. Hingham and Doe v. University of Michigan. Adjacent AI employment cases exist over algorithmic layoff selection, monitoring and hiring screens, but those are different claims and folding them in would misrepresent the record.

Read that absence carefully, because it does not mean the thing is not happening. Reuters reported in July 2026 that the widely predicted wave of AI-related employment litigation has not materialised, and legal experts gave two reasons: workers often have little visibility into how AI systems were used against them, and many have signed arbitration agreements that resolve disputes privately and keep them out of public court records. US District Judge William Orrick summarised the evidentiary problem facing the plaintiffs before him in one line, saying “they were not in the rooms where it happened.” That reporting concerns AI-driven layoff selection rather than authorship accusations, so extending its explanation to our narrower category is my inference and not a claim the article makes. It is still a documented reason why public court records would undercount this category however often it occurs.

In practice, for an employee, the work is unglamorous. Ask for the written policy the decision is being made under, which tool produced the reading, what number it returned, and which passages were flagged. Keep exchanges in email or a channel that retains history, and follow any meeting with a written summary. If the outcome starts moving toward termination, stop managing it alone and take the record you have been building to an employment attorney in your state.

Four marketplaces, four different answers. Most guidance flattens this into one sentence about platform policies, and the flattening is the error, because what your platform’s documents say determines which argument is even available to you.

Fiverr is the only one of the four with a real, on-point test. Its guidelines for freelancers and clients set out four things weighed when a client disputes a delivery as AI-generated: whether the client clearly communicated expectations about AI use before or at the start of the order; whether the delivery is accurate and free from clear mistakes or, in Fiverr’s phrasing, “AI-generated hallucinations”; whether it “reflects the freelancer’s own professional effort and meaningful input”; and, the criterion almost nobody quotes, “Whether both parties communicated professionally and responsively” throughout the order. Two things follow. If your client never stated an AI preference before the order started, the first factor is yours. The fourth is a reason to keep your replies measured even when the accusation stings. Consequences run both ways: Fiverr states that where it identifies a trust breach, including misrepresentation of AI usage, “the order may be cancelled, the client may receive a full refund, and the freelancer’s account may be permanently suspended.”

Those criteria sit in a guidance article, not in the dispute machinery. Fiverr’s Resolution Center, the actual mechanism, has no AI-specific provision: it is a request-and-response tool for cancellations, refunds and delivery extensions in which the other party has 48 hours to accept or decline, and, in Fiverr’s own words, “If the request submitted in the Resolution Center isn’t accepted or declined within 48 hours, the system will automatically accept it.” Read that twice if a cancellation request is sitting unopened in your inbox. Silence is not neutral there.

Upwork has an AI clause, and it is not about this. Section 2.4 of its User Agreement is a data licence: unless you opt out, you grant Upwork rights to use your platform content “to improve AI models exclusively provided to personalize your Upwork experience.” It governs training on your material, not a client’s claim that your delivery was machine-written, and nothing in the agreement addresses that claim. Worse for you, Upwork’s dispute-process page lists what it will not review, and the list includes “Subjective feedback about work quality or quantity” along with “Creative disagreements not backed by documented scope.” An AI accusation with no documented scope violation behind it may therefore not qualify for the formal process at all, which makes your contract terms, the brief and the written record the operative material rather than the dispute form.

WriterAccess has nothing. Its Terms of Use, read in full and searched term by term, contain no mention of artificial intelligence, generative AI or ChatGPT anywhere in the operative text. The single relevant string is “originality verification,” appearing inside a list of ranking criteria alongside performance ranking, star rating, reading level and language. Nothing defines what fails that verification, what evidence is weighed, or what a writer may do about it.

That absence is not academic. Gizmodo reported in June 2024 on Kimberly Gasuras, a news reporter of 24 years in Bucyrus, Ohio, who freelanced through the same marketplace, which Gizmodo spells WritersAccess. A detector called Originality flagged her writing. She received exactly one warning, wrote back to defend herself, and got no reply. Months later her account was suspended, and the message she was sent, as she described it, said the suspension was “due to excessive use of AI.” A platform with no published AI standard enforced one anyway.

Contently does publish an AI page, and calling it no policy would be inaccurate, so here is what it covers. Its AI governance page is about Contently’s own product: the main platform uses no AI, AI Studio is a separate opt-in product, and customer data is not touched by AI without a signed agreement. Searched for contributor, freelance, dispute and conduct, it returns nothing on any of them. It is a trust document for buyers of the platform, not a rulebook for the people writing on it. (Contently and Contentful are different companies with similar names, and a Contentful AI page surfaces in the same searches. They are not interchangeable.)

The instruction across all four is narrow and identical: before arguing anything, read your own platform’s current terms, work out which of these four situations you are in, and take the dispute to the mechanism that actually exists rather than the one you assumed existed.

Collect the work record that already exists. Everything here is preservation, never creation. Do not reconstruct a trail, backfill notes, or adjust a timestamp. Fabricated evidence is a far worse problem than the accusation you are answering, it tends to show up in file metadata, and it turns a survivable dispute into an unsurvivable one. Where a record does not exist, say so plainly and lean on the rest.

Paid work leaves a different trail from coursework, and usually a richer one. The brief or assignment thread shows what you were asked for and when. Scope and revision messages show the work being negotiated by a human. Interview notes, source emails, research links and the outline you sent for approval show the piece forming. Earlier accepted deliveries for the same client establish what your writing has always looked like to them, which persuades more than any general claim about your style. Delivery timestamps and the invoice put dates around all of it. Export or screenshot what you can, keep it in one folder, and keep the accusation thread beside it.

Then a caution, because overpromising here would repeat the error this page argues against. Process evidence is necessary and it is not always sufficient: Gizmodo’s 2024 reporting includes a copywriter who answered a 95% AI score by sending the client his drafting document with its edit-history timestamps, and lost the account regardless. Nobody outside a formal process is obliged to weigh what you send, which is exactly why the request for a human decision-maker in the first section is not a formality.

Alongside your own record, the published research on detector reliability is worth attaching, because it is the one element of your case that nobody has to take on trust. A 2023 study in the International Journal for Educational Integrity tested fourteen detection tools and reported that all of them “scored below 80% of accuracy and only 5 over 70%,” with accuracy falling further on machine-paraphrased text; note the date, since it describes 2023 tooling rather than today’s. That same year, researchers publishing in Patterns put authentic student essays written by non-native English speakers, with no model involved at any point, through seven detectors in common use, and measured an average false-positive rate of 61.3%, against 5.19% on a native-speaker control set. If English is not your first language, that gap is the most useful document in your file, and how second-language writing gets read by these tools explains why it exists rather than only that it does.

Our own numbers go here on the same terms, measured rather than asserted. We put human-written passages through our own detector and it holds up on exactly the writing this page is about: across 15,542 human-written passages it stayed clean on 99.8% (a 0.2% false-positive rate) and it flagged none of the non-native TOEFL essays or native-speaker student essays most at risk of a wrongful flag. The full breakdown is in our published measurements. No detector output is solid enough to decide a person’s income on its own (treat any score as a prompt to look closer, not a verdict), but a measured 0.2% false-positive rate is real evidence that a clean read on work you actually wrote means what it says.

Take the dispute to whichever process actually exists for you. Match the file you have just built to the mechanism from the previous section, because sending it into the wrong one spends your only leverage. On a Fiverr order it goes into the Resolution Center, and the 48-hour clock decides for you if you leave it. On Upwork, where a bare quality complaint may never be reviewed, it goes into the contract terms and the documented scope instead. On WriterAccess or with any direct client, there is no published route at all, so the agreement you signed and the person on the other end of it are the process. As an employee, it goes to the written policy and to HR. Work out which of those four you are in before you send anything, because the wrong door is how a good record ends up unread.

A candidate accused of using AI on a take-home task or written exercise holds the thinnest position of the three on this page, and I would rather say so than invent a process. Nothing published documents a candidate-side appeal route, because there is nothing to appeal to: no contract, no policy that covers you, no HR relationship, and no obligation on anyone to explain the decision. The other half is equally real. Nothing has been taken from you yet, and composure is worth more here than argument.

What you can reasonably do is offer to remove the doubt rather than dispute the number. Ask what specifically read as machine-written, since someone who can point at a passage is far more reachable than someone working from a bare percentage. Offer to do a comparable task live, or to talk the piece through in detail: how you framed it, what you cut, why it is built the way it is. That conversation is hard to fake and easy to have if the work is yours. Keep the tone level, because at this stage the accusation and your reaction to it are read together.

The live version is different and it is worth rehearsing once. If it happens in the room, over how you are speaking rather than over a document, because an answer sounded rehearsed or a pause ran long, say plainly that you are not using anything, then offer to take the question again in your own words. Keep it short. A long defence is the thing that reads badly, since the interviewer is judging composure as much as content, and an answer restated calmly settles the question faster than an explanation of why the suspicion is unfair.

Some employers will not reconsider, and knowing that in advance saves you a week. If you write in a second language, or your professional style is very clean and formal, the research in the previous section is context to mention once, briefly, without turning it into a grievance. Then move to the next process. There is no fair fight available here, only a good impression or a lost one.

If none of this has happened to you and you are reading as preparation, a separate guide covers pre-delivery checks on client work. It assumes nothing has been alleged, and it is the wrong page for a dispute already in progress, because at that point altering the delivered file works directly against you. Read it before a dispute exists, not after.