Three weeks into a search, the assistant stops being a novelty and becomes infrastructure. In week one the requests are cosmetic: tighten this bullet, fix the tense in the summary line. By week three it is rewriting that summary line for a fintech posting, drafting the follow-up to a recruiter who went quiet, and talking you through what to say when someone asks why you left after barely a year. By the time an offer is on the table it has seen the number. Nobody picks a tool for that trajectory, because nobody can see the trajectory on day one. It is whichever tab was already open. Anyone who moves to a ChatGPT alternative later, on privacy grounds, does it with a full history already behind them, which is a smaller improvement than it looks.
The reason is that the text moves faster than the reading does. A job search produces the most sensitive writing most people do in a given year, and it produces it in single sittings, under deadline pressure, in the same thread as a request to reformat a bullet point. The clause that governs what happens to it afterwards was written before the account existed and will read the same next month whether or not anyone opens it. That gap is the whole problem. By the time a search reaches the material worth protecting, the account holding it is weeks old, the history is long, and the decision about what the provider may keep was made by default, in a settings menu nobody opened.
What a job hunt actually puts in the box
The resume comes first. Home address, phone number, full employment history, and, on an older template, occasionally a date of birth. Then the drafts that never reach anyone: the honest version of why the last role ended, the note about a manager, the paragraph explaining a gap that involved a diagnosis. Then interview prep, which for a technical hire routinely means pasting in the current employer’s architecture so the assistant can help you talk around it. By offer stage, the conversation history holds compensation numbers, the name of the competing offer and your ceiling.
Regulators have noticed the incentive behind all this. The FTC has warned that model providers face liability when they break promises about customer data, including promises not to use it “for secret purposes, such as to train or update their models”, and specifically flagged companies that quietly rewrite terms of service to widen what they may keep. The agency’s position on privacy and confidentiality commitments is that there is no AI exemption from existing law. That is enforcement after the fact, and it does not retrieve a paragraph already absorbed into a training corpus.
In New York, the tool screening you has to be audited
Candidates in this city sit on both sides of the automation. Local Law 144 covers automated employment decision tools used by employers hiring in New York City. A tool must have passed a bias audit within the year before it is used, a summary of that audit has to be publicly posted, and candidates must receive notice at least 10 business days before it screens them. The Department of Consumer and Worker Protection began enforcing the law on 5 July 2023.
Read the notice when it arrives, because it tells you what the system evaluates. It also sets a reasonable standard for your own side of the exchange. An employer running software on your application owes you a published audit and advance warning. The assistant you run on your own application owes you the same clarity about what it does with the text afterwards, and rather more of it, since you feed that one voluntarily.
Settle the tool question before the offer stage
The compensation conversation lands in the same thread as the resume formatting request from six weeks earlier. Nothing about the account changed in between. What changed is that a searchable history now contains your salary floor and the name of a company you have not told your current employer about. A long search in this city compounds that, since the same agency recruiters work both sides of a vertical and the history keeps growing while you move between them.
Deleting the thread afterwards helps less than it feels like it should. A deletion in the interface removes what you can see. Whether it removes the backup copy, the moderation sample and the row already folded into a training set is a question the terms answer, and the answer differs by product.
Discipline is not the fix here, because what you would be disciplined about is invisible from the interface. Check the training policy in week one, when the search is still boring, and pick something whose answer you can live with in week ten. The hard conversations arrive on a schedule set by other people.
