What should never be automated in workplace misconduct management?

Building Ungender and Conduct means taking seriously both the difficulty of managing workplace misconduct and the limits of what technology should do within it. I want the process to be easier to navigate. I do not want that convenience to depend on people surrendering control over their accounts, or committees accepting conclusions they have not independently reached.

That is the purpose I want our work to serve: less avoidable administrative burden, more room for people to understand what is happening and exercise judgment. It is a standard to hold our work to, not a claim that a product can guarantee fairness.

The boundary cannot sit only at the final decision. By then, a complaint may have been rewritten, a respondent’s uncertainty removed, a bot’s reassurance mistaken for support and a summary accepted as the evidence. We need to examine the whole process.

The situations below are hypothetical. They are tests I would use when building or choosing a system, not accounts of particular people or cases. My proposed safeguards are design and practice judgments, not a claim that Indian law imposes a blanket prohibition on AI.

Receiving a concern without deciding what it means

Imagine an employee opening a reporting form after several uncomfortable interactions with a manager. The form requires a choice between sexual harassment, bullying and a grievance. The employee is unsure and wants to speak to someone first.

I would want that uncertainty to be an acceptable starting point. A person should not need to identify the correct legal category before reaching someone who can help. Routing is useful; an automated rejection that leaves no human route is a decision about access.

The receiving system should preserve what the person actually submitted, including tentative language. It should retain attachments and relevant submission metadata, keep later edits distinguishable, and explain who can see the information. Consent to optional AI assistance should be informed and separate from the ability to report. A person should be able to decline it without losing access to help.

There is also a legal reason not to treat writing assistance as an optional courtesy in every setting. Under section 9(1) of India’s PoSH Act, where an aggrieved woman cannot make the complaint in writing, the specified committee officeholders or members must provide reasonable assistance to do so.[4] That duty does not tell us to replace the person helping her with a generative tool.

Assistance should not manufacture a fuller complaint

Suppose the employee writes: “My manager repeatedly asked me to stay back after work, and I felt uncomfortable.”

An AI expansion might add that the manager exploited authority, threatened career progression or acted with sexual intent. Those matters might require examination. They are not details the tool is entitled to supply.

NIST’s Generative AI Profile describes confabulation as confidently presented false or erroneous content, including output that diverges from the input.[1] In a complaint process, my concern is that such additions could acquire the apparent authority of the complainant’s own words.

Translation, accessibility assistance and grammar correction deserve a different approach from invention. I would welcome assistance that helps a person express what they mean, provided they can check it in a language and format they understand. A polished account should not attract suspicion merely because someone needed help writing it.

The review needs to cover meaning. “I think this happened after the team dinner” must not become a definite date. A suggested chronology can sit beside the original; it should not silently replace the order in which events were recalled. A translated expression of discomfort should not become a stronger allegation without the person’s explanation and agreement.

I would preserve the original, show substantive changes and invite corrections without presenting the generated version as superior. Where information is missing, ask an open question. “What happened next?” leaves room for an answer. “How did the manager retaliate?” supplies a conclusion before the person speaks.

Both parties may arrive with AI-assisted accounts

The organisation’s system is only part of this discussion. A complainant might use a personal tool to prepare a statement. A respondent might use one to understand a long notice or draft a reply.

I would not regard either use as evidence of dishonesty. Someone facing an allegation also needs a fair opportunity to explain their account, and difficulty writing should not become a disadvantage. The same boundary applies to both parties: assistance with expression must not become invention of facts or substitution for their own recollection.

Consider a hypothetical respondent who writes, “I remember the meeting, but not the exact words.” A draft turns this into “I categorically deny making those statements.” That is a substantive change. Equally, a tool must not insert an apology that amounts to an admission the respondent did not intend to make.

A request to “make my case stronger” needs particular care. It may mean organising relevant material clearly. If it produces strategic omissions, invented corroboration or a more certain account than the person can support, the committee needs to examine those substantive problems, not the sophistication of the prose.

I would also worry about repeated exposure to suggested wording. Imagine a tool supplies a phrase supposedly spoken during a meeting. After several revisions, the person uses that phrase without remembering that it began as a suggestion. That hypothetical illustrates why I would keep independent recollection separate from generated language, rather than assume a final confirmation repairs everything.

For the committee, the useful inquiry is what the person can attest to and what supports it. I would not use writing style, AI use or an AI detector score to decide credibility. Neutral clarification and access to original material are more appropriate than asking either party to prove that they wrote unaided.

Immediate support needs someone who can act

An acknowledgement can tell a person their submission arrived. It cannot establish that someone has assessed what they need.

Imagine a complainant has a meeting with the respondent the following morning. A bot replies that the organisation takes safety seriously. The employee still needs to know whom to contact about that meeting, whether anyone has seen the request and what to do if the usual contact is unavailable.

I would allow a bot to explain an approved process, help locate a contact or arrange a request for assistance. It should identify itself as automated and distinguish a request logged from a handover accepted by a person. The organisation needs an actual escalation route, with ownership and a fallback, rather than a button that creates an unattended ticket.

A bot should not promise safety, absolute confidentiality or a particular outcome. It should not decide that a situation is too minor for urgent help, diagnose distress or improvise counselling. For urgent danger, it should make verified emergency and human contact routes visible without requiring the person to complete an AI assessment first.

The same care applies to assistance offered to respondents or witnesses. A reassuring conversation does not substitute for someone with authority to address a practical concern. Warm wording is welcome only when it accurately describes the support available.

Drafting an organisational response is consequential work

Now consider an HR colleague asking AI to prepare an acknowledgement. The generated email says the complaint has been reviewed, necessary safeguards are in place and the matter will be resolved promptly. In this hypothetical, nobody has yet reviewed it.

Each sentence sounds helpful. Each makes a claim the organisation has not established.

I would separate receipt, assessment and action in every reply. Before sending, the responsible person should check the actual status, dates, next step and contact route. An acknowledgement should not imply that the allegations have been accepted or rejected. Nor should a request for clarification introduce a fresh allegation on the complainant’s behalf.

A draft response to the respondent needs the same discipline. It should accurately communicate what requires a response, without substituting an AI interpretation for the account received or adding admissions and denials on anyone’s behalf.

Checking recipients is part of reviewing the substance. A correct message sent with the wrong attachment, copied to an unnecessary colleague or delivered through an unsafe contact channel can create another problem. I would want the sender to check access and the person’s stated communication preferences, not merely approve the tone.

Templates can help with consistency. Generative drafting should still leave a named person responsible for every factual representation the organisation sends.

A summary is not the record, and a report is not a prediction

An Internal Committee member might reasonably want help navigating messages, statements and dates. A draft chronology or an index could be useful. I would require each entry to lead back to its source, with allegations, responses, evidence and the tool’s interpretation kept distinct.

The original evidence should remain unchanged. Working copies, transcripts, translations and summaries should be identifiable as such, with corrections recorded. A reviewer should be able to establish where a statement came from and whether a quotation is exact.

Suppose a witness first gives an approximate date, then checks a calendar and corrects it. A system flags an inconsistency. That flag invites a question; it does not establish deception. The committee must hear the explanation and decide whether the difference matters.

NIST also identifies automation bias: people may defer excessively to automated systems or regard their output as better than other sources.[1] That risk is why I would not accept a generated credibility score or an eloquent recommendation as the starting assumption everyone else must disprove.

Tools should not predict the outcome a committee or employer would prefer and organise the reasoning around it. If drafting assistance is used for a report, it should work from findings and reasons the responsible committee has actually reached. It must not manufacture missing reasoning, citations or certainty. NIST specifically warns that generated logic and citations can themselves be false.[1]

A genuine reviewer needs the source material, time to examine it and authority to reject the output, seek clarification or stop the next step. “Human reviewed” means little if the person sees only a summary and an approval button. I would expect the decision makers to explain their reasoning without leaning on the tool’s confidence.

Confidential records should not become casual prompts

The convenience of pasting an entire case file into a public AI service deserves scrutiny before it becomes a habit. Removing names is not the only question; a detailed combination of role, workplace and incident may still make people recognisable.

NIST identifies privacy risks including exposure and inference of sensitive personal information.[1] But it would be inaccurate to say that every AI service trains on everything uploaded. Provider terms distinguish products and settings: Anthropic’s published policies, for example, separate consumer training choices from commercial services and describe different retention arrangements and exceptions.[2][3]

I would ask what the actual deployment permits: what is retained, whether training use is enabled, who can access the data, which connected services receive it and what deletion means. An enterprise label alone would not answer those questions. Neither would one party’s consent settle the handling of everyone else’s information in the file.

For PoSH matters, section 16 restricts publication or communication to the public, press and media of specified complaint and inquiry information, subject to its stated exception.[4] That is not a ready-made answer to every AI deployment question. Confidentiality, applicable obligations and the proposed data flow still need specific review before records are uploaded.

My practical starting point would be to keep identifiable case material out of unapproved tools and use only the information necessary in a reviewed environment. Access controls should apply to generated outputs too, not just the documents that produced them.

Responsibility should remain visible after the report

The workflow does not end when a document is uploaded. Someone needs to check what was communicated, what action remains and whether an agreed measure was implemented. A dashboard should not close a matter on behalf of a person who has not made that assessment.

This is the standard I want to bring to building Ungender and Conduct. Technology should help people keep track of the work while leaving them able to question it. The complainant should be able to correct an account. The respondent should be able to challenge an interpretation. A committee member should be able to reject a summary without the workflow pushing the same conclusion back at them.

Before approving a tool, I would take a hypothetical matter through every stage, including an uncertain account, an urgent support request, a mistaken draft and a disputed finding. At each point, I would want to locate the person who can explain what happened and put it right. If that person becomes harder to find as the process becomes faster, we have more work to do.

Sources

[1] https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf — NIST AI 600-1: Generative AI Profile (July 2024)

[2] https://privacy.anthropic.com/en/articles/7996866-how-long-do-you-store-my-organization-s-data — retention

[3] https://www.anthropic.com/news/updates-to-our-consumer-terms?subjects=claude — training

[4] https://doe.gov.in/files/inline-documents/DoE_Prevention_sexual_harassment.pdf — PoSH Act 2013, official Department of Expenditure copy

Penned by Pallavi Pareek, Founder of Ungender and Conduct.