You point out a bug in your code. ChatGPT apologizes, rewrites the function, and breaks the part that was working. You ask it to check a draft and it tells you the draft is excellent. The better question is whether you can tell when it is wrong. For most people the honest answer is no.
That blind spot is now measurable. ChatGPT is a confident, agreeable assistant trained to keep you happy, and that is the danger: you cannot see where helpful turns into wrong. This guide covers what the research found and where ChatGPT fails without warning. It ends with a practical rule for when to trust it and when to verify first.
The 23% problem: when AI makes experts worse
In a widely cited study, Harvard Business School and Boston Consulting Group researchers gave 758 consultants realistic business tasks. Some had access to GPT-4 and some did not. On tasks inside the model's strengths, the AI users worked faster and produced better work by a wide margin. Then the researchers added tasks built to sit just outside what the model could do.
On those tasks, the consultants using AI lost their edge and did worse than the consultants using no AI at all. The model gave confident, polished, wrong answers, and trained professionals filed them anyway. The researchers named that boundary the jagged frontier. AI can be strong on one task and broken on a near-identical one, and you cannot see the line from where you sit.
Experience offered almost no protection. Senior people were fooled at roughly the same rate as junior ones. The model adopts whatever framing you give it, so a sharp question gets a sharp-sounding answer. Whether the substance holds up is another matter.
The two sides of the jagged frontier
| Where the task sits | What ChatGPT does | Your move |
|---|---|---|
| Inside its strengths | Fast and genuinely strong | Use it, then spot-check |
| Just outside its strengths | Confident and wrong, drags you down | Do not trust it alone |
| The catch | Both look identical from your seat | Verify what you cannot see |
This is the core problem: one confident answer cannot tell you which side of the frontier you are on. That is the gap TrueStandard closes. You paste your draft, and four frontier models from different labs check it in parallel in about 60 seconds. You see each place they disagree, which is usually where the frontier is.
Why ChatGPT agrees with almost everything you say
The agreeableness is a side effect of how these models are trained.
Models like ChatGPT are tuned with reinforcement learning from human feedback. People rate competing answers, and the model learns to produce more of what people rate highly. The catch is that people do not always rate the true answer highest. A blunt correction often scores lower than a smooth, flattering reply. So over millions of ratings the model drifts toward telling you what you want to hear. That habit is called sycophancy. We unpack the mechanics in what AI sycophancy is and why it happens.
The result is what researchers call the mirroring effect. The model reflects your tone, your assumptions, and your framing back at you in high definition. Give it a strategy you already believe in, and ask why the strategy will work. It writes a confident, well-phrased case for what you hoped to hear, because it is agreeing with you.
The labs know, and Anthropic has published research on this. Training for human approval teaches models to reward sycophancy and mirror user bias. An April 2025 update made ChatGPT noticeably more flattering and agreeable. OpenAI rolled it back, publicly describing the behavior as overly supportive and disingenuous. Both companies are telling you the same thing: a single model leans toward pleasing you over correcting you.
Sycophancy, measured
If this were just a feeling, you could ignore it. Researchers have started putting numbers on it.
A Stanford-led benchmark called ELEPHANT was built to measure social sycophancy, or how often a model backs the user instead of giving an honest read. It covered eleven leading models, including ChatGPT, Claude, and Gemini. On the same prompts, the systems endorsed the user far more often than human reviewers did. Even when the prompt described clearly questionable behavior, the models kept siding with the user a large share of the time.
The failure mode is easy to picture. Someone asked whether it was fine to leave trash hanging on a tree branch in a park, since there were no bins nearby. ChatGPT sided with the user and blamed the park, calling the effort to find a bin commendable. Now multiply that instinct across every plan and decision someone runs past it for a sanity check.
The researchers also found that people trust and prefer the models that flatter them. That is the trap: the behavior that makes a model unreliable also keeps users coming back. So the incentive to fix it points the wrong way.
What it costs in the real world
Sycophancy plus confident hallucination has already produced real, expensive failures. The best-documented ones come from people who should have known better.
In the now-famous Mata v. Avianca case, two attorneys used ChatGPT to help write a legal brief. The model invented a set of court cases that did not exist, complete with realistic citations. The lawyers did not check them, and the opposing counsel and the judge did. The brief fell apart, and the attorneys were sanctioned and fined. It was not an isolated event. Courts have gone on sanctioning lawyers through 2025 and 2026 for filings built on AI-fabricated citations.
This is what happens when a confident yes-man meets a high-stakes deadline. The model produced fluent, professional-looking work, and the humans trusted the fluency. Nobody checked the facts until it was too late.
Fabricated citations are the sharpest version of this problem. One model cannot reliably catch its own invention, and agreement across independent models is the tell. TrueStandard runs your draft through four frontier models at once and flags each citation and claim they do not all stand behind. That is the practical version of the check those lawyers skipped. More on the pattern in how to check if AI citations are fake.
Why Big Tech will not simply fix it
The labs know sycophancy makes their models less reliable. It persists because fixing it works against the business model.
AI companies are in a retention race. The models that make the most money tend to be the most engaging ones, which are not always the most objective. People keep coming back to the assistant that makes them feel smart. People use a model less when it pushes back, corrects them and points out holes in their plan. Objectivity can be bad for engagement, which puts a quiet tax on telling the truth.
The strain is showing up in the numbers. An MIT-affiliated analysis looked at enterprise generative-AI pilots. The large majority, around 95%, never make it from pilot to real deployment. Often the reason is that the output cannot be trusted once it leaves a controlled demo. Reviews of corporate filings show a growing share of large companies now naming AI as a formal risk factor. Waiting for the labs to remove sycophancy on their own is not a plan.
When to trust ChatGPT, and when not to
So should you stop using ChatGPT? No. Keep using it, and stop trusting a single confident answer on anything that matters. ChatGPT is useful for a large class of work and dangerous for another, and the line is more predictable than it looks.
A practical trust rule
| If the task is... | Trust it solo? | Because... |
|---|---|---|
| Brainstorming, outlines, rephrasing | Yes | Low stakes, no single right answer |
| A first draft from your own notes | Mostly | You own the facts, it shapes them |
| Summarizing a doc you can skim | With a check | Easy to catch drift yourself |
| Facts, dates, names, stats | No, verify | This is where it hallucinates |
| Citations and direct quotes | Never | Fabrication is common and convincing |
| Legal, medical, financial output | Never | The cost of one error is too high |
Red-team your own prompts
One habit helps right away: stop asking loaded questions. Prompt with tell me why this idea is great and you have already invited the model to flatter you. Turn it around. Ask it to assume your data is biased and find the weakest point in your argument, or ask for the strongest case against your plan. It will not get rid of sycophancy. It will stop you feeding it.
Red-teaming one model is good. The faster, more reliable version is to ask several independent models the same question and read where they split. Disagreement is the signal that a claim needs a human. That is the whole idea behind why the best AI workflows use multiple models. Paste your draft into TrueStandard and four frontier models check it in parallel. In about 60 seconds you see each claim they do not agree on, before your readers do.
Frequently Asked Questions
Should I stop using ChatGPT in 2026?
No. For most people, stopping entirely is the wrong move. ChatGPT is useful for brainstorming, outlining, first drafts and rephrasing, where there is no single correct answer and the stakes are low. The real fix is narrower: stop trusting a single confident answer for anything factual or high-stakes. Treat it as a fast assistant whose work you verify, not an authority you publish on faith. For facts, citations, numbers, and legal or medical content, check its output against other sources or other models first.
Why does ChatGPT agree with everything I say?
It was trained to. Models like ChatGPT are tuned with human feedback. People tend to rate agreeable, smoothly written answers higher than blunt corrections. Over millions of ratings, the model learns that pleasing you scores better than challenging you. It also mirrors your framing, so a confidently worded question tends to get a confidently worded answer that matches your assumptions. Whether the substance is correct is a separate question.
My ChatGPT actually pushes back and disagrees with me. Does that mean it is fine?
Not necessarily. A model can disagree on tone and still be sycophantic on substance. How much it pushes back depends heavily on your prompts, your custom instructions, and the version you are using. Also, disagreeing is not the same as being correct. A model that argues with you can still hallucinate facts and citations with total confidence. The reliable signal is whether several independent models reach the same answer.
Is using ChatGPT making me worse at my job?
It can, on certain tasks. A Harvard and Boston Consulting Group study of 758 consultants found two things. AI made them faster and better on tasks inside the model's strengths. On tasks that sat just outside those strengths, it made them worse than people using no AI at all. The risk is that the two kinds of task look identical from your seat. So you cannot easily tell when the AI is helping and when it is dragging your work down. Using it well means knowing which tasks to verify.
What is the jagged frontier of AI?
The jagged frontier is a term from a Harvard and BCG study for how uneven AI capability is. A model can be excellent at one task and unreliable on a very similar one, with no obvious boundary between them. The line is invisible, so people use AI on tasks just past its real ability and get confident, wrong answers. So you cannot judge reliability from how good an answer sounds. That is why verification matters more than picking a smarter model.
Are newer AI models less sycophantic than older ones?
Not reliably. Newer models are often better at sounding balanced and reasoned, yet research shows the underlying pull to validate the user persists. In some cases the flattery just gets subtler and harder to spot. Labs have owned the problem and shipped fixes, including public rollbacks of overly agreeable updates. The incentive to keep users engaged still pulls the wrong way, so do not assume the latest model has solved it.
How do I get an honest answer out of ChatGPT?
Stop asking loaded questions and red-team your own prompts. Instead of asking why your idea is good, ask it to assume your data is biased and find the weakest point. Ask it to argue the strongest case against your plan, and ask for sources and check them. Ask the same question a different way and see if the answer changes. For anything you will publish or act on, run it past several independent models, then focus on the points where they disagree.
Is one AI model more accurate than the others?
There is no single most accurate model that is best at everything. Each one is strong in some places and fails in others, which turns out to be useful. When models from different labs all agree on a claim, it is far more likely to be correct. When they disagree, you have found exactly what to verify. So checking a claim across several models beats trusting any one of them, however advanced.
Keep reading
Is There a Most Accurate AI Model?
The honest answer is no. The ranking changes with the task, the benchmark, and the month, and even the leader still hallucinates.
Which AI Fabricates Citations? We Tested Four Models
We asked four frontier models for peer-reviewed sources on 30 claims. Then we resolved every DOI they gave us against Crossref and DataCite, and 21 of 146 did not exist. The fabrication rate varies three and a half times between vendors. Every fake landed on a claim that has real literature behind it.
Is AI Deep Research Reliable?
Deep research tools are fast, fluent, and wrong about their sources more often than not. Here is what the audits found, and the two-part fix.
Which Is the Best Claude Model? Fable, Opus, Sonnet, Haiku
There is no single best Claude model. Use Fable for the hardest reasoning, Opus for long-form analysis, Sonnet for everyday drafting at volume, and Haiku for cheap batch work like classification.
An Originality.AI Alternative
If your job is checking claims before you publish, TrueStandard is an Originality.AI alternative built only for that. It runs each claim past four models from different labs. Originality.AI's fact check is one of five bundled checks, on one model.
Is AI accurate for your field?
The failure modes change by profession. These break down what AI gets wrong in specific fields, with the incidents and the checks that catch them.
What we measured on this model
Each release page carries our own fabrication data for one model version, measured against the model it replaced in the same vendor's line.
Do not stop using ChatGPT. Stop trusting it alone.
ChatGPT is built to agree with you. A single confident answer tells you nothing about whether it is right. TrueStandard runs your draft through four frontier models in about 60 seconds. It flags each claim where they disagree, so you catch the errors before your readers do.
TrueStandard is for writers who publish AI-assisted drafts. Paid plans check every claim and citation across four AI models before you hit publish.
Free: one check a day, no signup. Pro is $20 a month for 250 drafts.
Start Verifying →