Job search · Updated 11 September 2026

How to Use AI to Negotiate Your First Job Offer

Use it for the words, the rehearsal and the reading of the offer letter. Do not use it for the number.

Ask a chatbot to draft your counter-offer email and it will do a good job. Ask it what number to put in that email and it will invent one, confidently, from nothing. The split matters because the number is the only part of a negotiation that is actually about the world: it comes from posted pay ranges, your university's own graduate outcomes report and government wage data, all of which are free and none of which a language model has in front of it. Get the figure from those, then let AI handle the wording, the questions you forgot to ask, and the practice run.

Why the chatbot's number is not an estimate

This has been measured properly, and the results are worse than "it might be a bit off".

In February 2025, PLoS One published a controlled audit by R. Stuart Geiger, Flynn O'Sullivan, Elsie Wang and Jonathan Lo with the unusually direct title "Asking an AI for salary negotiation advice is a matter of concern". They ran bias audits on four versions of ChatGPT, asking each "to recommend an opening offer in salary negotiations for a new hire", submitting 98,800 prompts to each version while varying the candidate's gender, university and major.

Their finding: "ChatGPT as a multi-model platform is not robust and consistent enough to be trusted for such a task." Two results explain why.

• The biggest gap was between model versions, not between candidates. The newest model tested recommended roughly 40% higher salaries than the oldest — a median difference of about $25,000 for the same candidate asking the same question. Which chatbot you happened to open moved the answer by more than any fact about you.
• The answer flipped with whose side you asked from. Prompts written in the employee's voice produced recommendations $7,500 to $25,000 higher than the same scenario asked in the employer's voice. The model is not computing a market rate. It is producing the kind of number the person asking seems to want.

Who you are moved it too. Computer science majors were recommended about $6,000 more than the control, humanities and social science majors $8,000 to $12,000 less, and graduates of elite coastal private universities were advantaged by $4,200 to $8,100. The researchers also tested fictional and fraudulent universities and reported "wildly inconsistent results across different cases and model versions" — the model was rewarding the sound of a university name.

A second study, "Surface Fairness, Deep Bias" by Aleksandra Sorokovikova, Pavel Chizhov, Iuliia Eremenko and Ivan P. Yamshchikov, presented at the 6th Workshop on Gender Bias in NLP in Vienna in August 2025, widened this beyond one vendor. Testing Claude 3.5 Haiku, GPT-4o Mini, Qwen 2.5 Plus, Mixtral 8x22B and Llama 3.1 8B, the authors report that models score personae near-identically on a knowledge benchmark, but: "if we ask the model for salary negotiation advice, we see pronounced bias in the answers." Comparing their extreme profiles, a "Male Asian expatriate" persona was recommended more than a "Female Hispanic refugee" persona in 87.5% of tested scenarios.

Where the number actually comes from

All four of these are free, public, and specific in ways a chatbot cannot be.

The posted range

Pay transparency rules have quietly done most of the research for you. Washington State's Equal Pay and Opportunities Act, for example, requires employers with 15 or more employees to include in the posting "a wage scale or salary range" along with "a general description of all benefits" and "a general description of other compensation". More than a dozen US states and several cities now require some form of disclosure, but the triggers differ — some in the advert, some only on the applicant's request, some by the time an offer is made — so check your own state labour department rather than assuming. If nothing is posted, asking the recruiter for the band is a normal question, not a rude one.

Your university's outcomes report

This is the source almost no negotiation guide mentions, and it is the closest thing you have to data about people exactly like you. Most universities run a first-destination survey, a national standard collected by NACE that records what graduates are doing within six months of finishing, including starting salary. Many publish it as a browsable dashboard: Virginia Tech's invites you to "Explore post-graduation outcomes shared by our graduating seniors" and notes that "the salary metrics listed are based on student-reported starting salaries and do not include bonuses or other possible monetary benefits". UConn reported an average starting salary for its Class of 2025 of "$67,500, $3,500 higher than last year".

Search your own career centre's site for "first destination" or "career outcomes". A median for your major at your institution beats any national average, and it is a citable number in a negotiation.

Government and national data

The Bureau of Labor Statistics publishes Occupational Employment and Wage Statistics, currently reflecting May 2025, with estimates by occupation at national, state and metropolitan-area level — so you can look up what your job title pays in your city rather than in the abstract. NACE's Class of 2026 projections put computer sciences highest at an average of $81,535 and engineering at $81,198, with most categories rising 3.1% to 6.9% and social sciences projected to fall 1.7%. Treat those as employer projections for broad categories, not an offer you should expect.

What AI is genuinely good for here

Four jobs, none of which require it to know a number.

1. Read the offer properly. First offers are the first legal document most graduates receive, and base salary is the part people fixate on precisely because it is the part they understand.

Below are the terms of a job offer. I have removed my name and address.

Do NOT tell me whether the pay is good - you have no way to know that.

1. List every term of compensation and benefit stated here, in plain
   English, including anything conditional.
2. List what a full offer of this kind normally specifies that is
   MISSING here.
3. Give me the questions I should email the recruiter to close those
   gaps, phrased neutrally.

[paste the terms]

2. Expose its own instability. This is the study turned into a two-minute demonstration. Open two fresh chats and ask the identical scenario twice — once as the candidate, once as the hiring manager.

Chat A: I am a hiring manager making an offer to a new graduate for
[role] in [city]. What opening base salary should I offer? Give one
specific number.

Chat B: I am a new graduate who has been offered [role] in [city].
What opening base salary should I ask for? Give one specific number.

Compare them. The gap between two answers to the same underlying question is the model's own estimate of nothing, and seeing it is more persuasive than being told. Use the exercise to stop trusting the figure, then go and get a real one.

3. Draft the counter using your numbers. The model supplies the sentences; you supply every digit.

Help me write a short, warm counter-offer email.

Use ONLY these numbers. Do not add, adjust or suggest any figure of
your own, and do not estimate what I am worth.

Role: [title]. Offer: [base]. My counter: [number].
Evidence I am citing: [posted range / my university median / BLS figure].

Rules: under 150 words. Accept the role enthusiastically first. One
clear ask with the evidence attached. No apologising, no
over-explaining, and end with a question rather than an ultimatum.

4. Rehearse the awkward middle. The email is easy. The call where someone says "that is above our band for this level" is the part that goes wrong.

Role-play a recruiter who has received my counter-offer. Be realistic
and slightly resistant, not hostile and not a pushover.

Push back at least twice - once on budget, once by asking me to justify
the number - and at some point ask what my minimum is.

Ask one question at a time and wait for my reply. Do not coach me
mid-conversation. At the end, tell me which of my answers were vague
and which gave away information I did not need to give away.

The offer is [base] for [role]. I am asking for [number].

The three mistakes the chatbot will encourage

• Naming a number before you have data. If a model hands you a figure in ten seconds, you will anchor on it. Look up the posted range and your school's median first, then open the chatbot.
• Negotiating only base salary. Ask the model to list what else is on the table — start date, signing bonus, relocation, review timing, holiday, the health premium — and you will usually find something that moves when base will not.
• Sounding like everyone else. A model produces the most likely phrasing, so the recruiter reading twenty counters this month has seen your email already. Write the first version yourself and let AI tighten it, the same discipline we set out in keeping your own voice when AI gives feedback.

Where this fits in the rest of the job search

• Everything before this stage. The offer is the last step of a funnel that starts with getting your resume past AI screening and runs through behavioural interviews and what employers actually permit you to use AI for.
• Check the offer is real first. An offer that arrives without a proper process, or that asks for money or bank details early, is not a negotiation problem — see how to spot an AI job scam.
• The habit underneath all of this. Treat model output as a list of things to confirm, never as the confirmation, which is the whole of our AI hallucination checklist.

FAQ

Should I just ask ChatGPT what salary to ask for?

No. A PLoS One audit of four ChatGPT versions found the largest gaps were between model versions, with the newest recommending roughly 40% higher salaries than the oldest tested, a median difference of about $25,000 on the same request. The same study found the model recommended different numbers depending on whether the question was asked in the voice of the employee or the employer. A figure that moves with the model you opened and the way you phrased the question is not an estimate of your market value.

Is it safe to paste my offer letter into a chatbot?

Not as it arrived. An offer letter usually carries your full name, home address and sometimes a national insurance or Social Security number, and most chatbots retain conversations by default. Retype or redact the terms you want help with — title, base, bonus, start date, benefits — and leave every identifier out.

What if no salary range is posted where I live?

Ask the recruiter directly for the band for the role, which is a normal question, and check your state or national labour department, because disclosure rules differ and some require the range on request or by the time an offer is made. Meanwhile you still have three free sources: your university first-destination report, the BLS Occupational Employment and Wage Statistics for your occupation and metro area, and the NACE starting-salary projections for your major.

Will asking for more make the employer withdraw the offer?

Nobody can promise it never happens, and anyone quoting you a precise probability is guessing. What you can control is the shape of the ask. A single polite counter, grounded in a published figure and paired with genuine enthusiasm for the role, is an ordinary part of hiring. Where a range is already published, a number inside that range is asking for something the employer has stated in public that it pays.

Bottom line

Split the task in two and the tool stops being dangerous. The number is a claim about the world, and it belongs to the posted range, your university's outcomes report and government wage data — sources that do not change their answer because you rephrased the question. Everything else in a negotiation is language and nerve, and that is where a model earns its place: reading the offer closely, finding the terms you forgot to ask about, tightening your email, and sitting across from you for the rehearsal so the real conversation is the second time you have had it.