An AI Summary Changes What Your File Has to Show

Does reading a summary first make the decision automated?

Almost certainly not, and the reason sits in one definition rather than in how much the summary influenced anyone. New York City's rules define a simplified output as a prediction or classification, which may take the form of a score, a tag or categorization, a recommendation, or a ranking ¹. A paragraph describing what happened in a session takes none of the four forms.

The same rules close the nearest gap. Simplified output, they say, does not refer to the output from analytical tools that translate or transcribe existing text, and the example given is transcribing a video or audio interview ¹.

So the phrase everybody quotes never gets reached. Substantially assists or replaces discretionary decision-making is a test applied to a simplified output, and with no simplified output there is nothing to apply it to. The three tests that decide whether a tool is an AEDT stop at the first element.

A session summary does not always stop at narration, though.

A summary ending in a rating, a label such as strong or partial, or a line recommending the candidate advance is a simplified output on the rules' own wording.

Then the three conditions decide it: relying on the output alone with nothing else considered, weighting it more heavily than any other criterion in the set, or using it to overrule conclusions derived from other factors including human decision-making ¹.

The duties that follow arrive as a package: a bias audit conducted within the prior year, a public summary of its results, and notice to the candidate at least ten business days before the tool is used. The law took effect on January 1, 2023, and enforcement has run since that July ².

The EU rule is aimed at the reviewer, not at the summary

The EU's answer looks past the summary at the person holding it. Providers of high-risk systems have to enable the people assigned to oversight to stay aware of the possible tendency to rely, or over-rely, on what the system produces, which Article 14 of the AI Act calls automation bias ³. The clause singles out systems that supply information or recommendations for decisions taken by natural persons.

Hiring is on the AI Act's high-risk list. Recruitment and selection are listed by name, along with the analysis and filtering of job applications and the evaluation of candidates (Annex III, point 4) ³.

The date is later than most compliance notes say. The European Commission's own implementation timeline puts the Annex III high-risk rules at 2 December 2027 ⁴. The regulation as adopted applied from 2 August 2026 ³, and a great deal of 2025 commentary still quotes that date.

The duty that lands on a hiring team is a staffing standard rather than a software standard.

Oversight has to be assigned to natural persons who hold the necessary competence, training and authority, and who are given the necessary support (Article 26(2)) ³.

A reviewer handed a summary, given no account of what it was built from, and holding no authority to disregard it falls short on authority and on support.

The older European rule turns on a single word. GDPR Article 22 gives people the right not to be subject to a decision based solely on automated processing that significantly affects them, and Recital 71 names e-recruiting practices without any human intervention as its example ⁵. The right has applied since May 2018, so none of it is new.

A rubber-stamp reviewer who never varies from the machine's output is not meaningful human intervention, though the article itself sets no threshold.

How far solely reaches was tested in a case about a credit agency rather than an employer (Case C-634/21). Establishing that agency's automated probability value about a person is itself automated individual decision-making, the Court of Justice of the EU held on 7 December 2023, where a third party receiving the value draws strongly on it ⁶.

No court has applied the reasoning to a candidate, and draws strongly on is a fact-specific phrase rather than a line. What travels is the structure: the thing producing the value can be the decision-maker even when somebody else formally decides.

Read the primary evidence first, then the summary

The order is the whole intervention, and it costs a few minutes. Have the reviewer open the candidate's actual work, the transcript, the submitted artifact, or the recording, and form a view before the summary is available to them. Then let them read the summary and record whether it changed anything.

A reviewer who reaches a score before seeing the summary leaves a file showing the first score came without it.

Record the session in the order it happened, so the file itself shows which came first:

• The primary evidence the reviewer opened, timestamped.
• The view they wrote down before the summary was available to them.
• The summary, stored exactly as the tool produced it, labeled with the tool and the version that wrote it.
• The score, plus one line on whether the summary changed it.

Write the reviewer's note in what the candidate did rather than in what the summary said about them. Which words in a screen note create a problem later is the same discipline applied one step earlier in the loop.

The summary is itself a document about a person, so whether an AI notetaker's summary can go in the hiring file turns on what it contains. A line reporting the answer a candidate gave is evidence. A line reporting that the candidate seemed nervous and rehearsed is a conclusion, and the file cannot support it.

No measurement of whether reading order changes a score in a hiring review sits behind this advice. Reading the evidence first is a design choice made because the downside is cheap, not a finding. A reviewer under deadline may well open the summary first unless the file makes the order visible.

Check the transcript the summary was written from

A summary of a session is a summary of a transcript, and the transcript is not neutral. Machine transcription carries measured error that falls unevenly across speakers, so two candidates who said the same thing can reach the reviewer having said different things. Ask the vendor what the summary is generated from, and whether a reviewer can open that layer in one click.

The gap was published in 2020, from systems tested in 2019. Five commercial speech-recognition systems transcribing matched sociolinguistic interviews with 42 white and 73 black speakers produced an average word error rate of 0.35 for black speakers against 0.19 for white speakers ⁷.

The study measured transcription accuracy only, on research interviews rather than job interviews, and its authors traced the gap to the acoustic model and to a shortage of training audio from black American speakers.

Roughly one word in three came out wrong for one group, and roughly one in five for the other.

An audit of one version of OpenAI's Whisper, run on a research speech corpus and not on job-interview recordings, found something worse than an error. About 1% of transcription segments contained entire hallucinated phrases or sentences appearing nowhere in the audio, and 38% of those carried explicit harms such as invented violence or implied false authority ⁸.

Four other commercial services produced no comparable hallucinations on the same segments, so the finding belongs to one model version rather than to speech-to-text as a category ⁸.

The hallucinations landed disproportionately on speakers with longer non-vocal stretches, a common symptom of aphasia ⁸.

A fabricated sentence is indistinguishable from a real one inside a transcript. A summary written on top of one inherits the fabrication and drops the audio that would have caught it.

A reviewer needs a route back to the recording rather than a better summary.

Write down whether the summary is a simplified output

One page, written now, answers whether the summary automates the decision, and it answers in the form a regulator or a candidate will want. Name the tool. State whether its output is a score, a tag, a recommendation or a ranking, state what the reviewer reads first, and name who holds authority to disregard the summary. Date the page, and date the next review of it.

Colorado supplies the deadline. Governor Polis signed SB 26-189 on 14 May 2026, repealing and reenacting the state's 2024 AI act as a narrower automated-decision-making regime whose requirements start on 1 January 2027 ⁹.

The replacement names employment a consequential decision, requires notice at the point of interaction, requires a plain-language description of the technology's role within 30 days of an adverse outcome, and gives a right to request meaningful human review ⁹.

A plain-language description of the technology's role is exactly the document a reviewer's summary makes harder to write.

No US law gives candidates a right to a human reviewer today; Colorado's replacement act adds a right to request one.

Read the replacement narrowly. Its summary of the new requirements carries no duty of care against algorithmic discrimination, enforcement sits with the attorney general alone, and a law amended twice in two years may be amended again ⁹.

None of this is legal advice, and dates in this area move; confirm current status with employment counsel.

The summary in the reviewer's hands was never the problem worth solving. Labeled, kept and read second, it is a note on a file that already holds the evidence. Read first and thrown away, it becomes the only account of the session anybody has, and the candidate cannot see it.

Decide which of the two your file holds. Colorado's clock starts on 1 January 2027.

Read the evidence >