We sent a summary of allegations to a complainant for them to review recently, which is a routine step and one of the more important aspects of the process, because a party deserves the chance to tell us we have misunderstood them before anyone else reads a word of it. Two pages, nine allegations, each set out in plain language with the particulars taken from the interview.
What came back was forty pages.
It was not a correction of our summary; it was a statement of the case. Sub-complaints nested inside the original complaints, incidents from years earlier that had never come up in the interview, exhibits appended and referenced by number, and argument, paragraph after paragraph of it, explaining why each incident satisfied the legal definition of harassment and what the employer’s exposure was likely to be. The tone bore no resemblance to the person interviewed, who had been nervous, plain-spoken, and quite compelling in their own words.
This Complainant was not being difficult, and that is an important thing to keep in mind. They had run our summary through an AI tool, asked how to make the complaint stronger, and it gave them exactly what they asked for. What it could not say was that forty pages would now cost them, and everyone else involved, weeks while we sorted out what was in scope, what amounted to fresh allegations, and how much of it the respondent could fairly be asked to answer. The Complainant had worked hard, in good faith, to make their own process longer and harder…and nobody had warned them.
We have been watching the wrong side of the table
Most of our conversation about artificial intelligence in this field has been about ourselves. Can we use it to transcribe an interview, to organize evidence, to help draft a report, and what do we owe the parties by way of disclosure if we do? Fair questions, and I have spent a good deal of time on them. But while we have been examining our own practices, the parties have quietly become the heaviest users of this technology in the process, and that will reshape our work far more than anything we choose to do with it ourselves.
The polished complaint that says less than it seems
The complaints themselves have changed too. They are cleaner and structured into numbered allegations, they cite the Occupational Health and Safety Act and the Human Rights Code, sometimes accurately and sometimes not very well at all, and every so often they reference decisions that do not exist.
Here is the irony…a better organized complaint is often a harder complaint to investigate. What makes an account useful to us is the presence of particulars, the small and awkward details that let us test it, such as who else was in the room, what was said just beforehand, and whether the person went back to their desk afterward and told someone. Those particulars are the first thing to go when a party asks a language model to make their story sound professional, because the machine is working toward polish while we are working toward evidence.
Credibility: the gap between the page and the person
This creates a real problem for credibility assessment with the risk running in two directions. We have always paid attention to how a person tells their story, looking at spontaneity, consistency, peripheral detail, and whether the account shifts under questioning. When a written complaint reads like counsel drafted it and the same person then speaks in an entirely different register, the gap is noticeable, and I suspect some investigators will quietly treat it as a credibility concern. Occasionally it may be one. Often it will not be, and this is where we need to be disciplined. The party who used AI to help write their complaint may be a second-language speaker who wanted to be taken seriously, or someone with a learning disability, or a person who feared that poor grammar would get their allegation dismissed out of hand. For those parties AI is an accessibility tool and a real step toward access to justice, and if we penalize them for using it, we will have harmed the very people the process exists to protect. Quite simply, we assess the person in front of us rather than the prose.
The coached interview
The same thing shows up in the interview room. Parties now arrive having asked a chatbot what they will be asked and how they ought to answer, and respondents in particular, come in with something close to a script. Ask what happened at the meeting and you will hear that the conduct was reasonable in the circumstances, was not unwelcome, and was certainly not vexatious. The vocabulary of the Code arrives before the facts do.
Preparation itself is not a bad thing, and representatives have been preparing parties for as long as this work has existed. What has changed is that it used to require a lawyer or a union representative and now it requires a phone and thirty seconds. I am glad of the levelling that brings, but it also means we are more often interviewing a prepared position rather than a person’s memory, which is why our craft has to adjust with more open questions, more patience, and more willingness to go sideways into the detail no model can anticipate for a party.
Volume, and who pays for it
The forty pages were not an isolated event. Objections to process, challenges to our independence, demands for documents, submissions responding to submissions, complaints about the investigator sent to the employer mid-file. None of it is new but generating it now costs a party almost nothing while reading and answering it properly still costs real time and money. Fairness requires that a party’s concerns be taken seriously even when we suspect they have been manufactured. The result is longer files, higher costs, and delay, which is what harms parties most. So, are we managing these processes, or has the volume begun to manage us?
Evidence we can no longer take at face value
I would also point out what our digital forensics colleagues have been warning us about for a while. A screenshot of a text exchange, or an audio clip of a manager saying something indefensible, can no longer be treated as self-authenticating, because fabricating a convincing screenshot was once beyond most people and is now trivial, and synthetic audio of a specific voice is no longer out of reach for anyone. That is no reason to treat every party as a suspected fabricator, which would be unfair, but provenance matters far more than it used to. Where did this come from, what is the original, and can we get the record from the source system instead of a party’s phone?
The party who arrives already certain
This last point concerns me most, because it is about people rather than process. Parties are talking to these tools before they ever talk to us, sometimes for weeks, and the tools are agreeable by design. Someone describes their situation, the model reflects it back sympathetically, confirms that what happened sounds like harassment, and sets out the remedies they can expect, occasionally drawing on American law that has no application here. By the time they reach us they hold a settled view of the merits and the outcome, arrived at where nothing they believed was ever tested. When the finding comes back unsubstantiated, and sometimes it will, they have much further to fall, and we will be accused of bias by someone who sincerely believes it.
Where I think we go from here
None of this is an argument for wishing the technology away. It is here, our parties are using it, and a good deal of that use is legitimate and even helpful. Of course we cannot control it, but we can be deliberate about how we respond.
- Say something at the outset: our process letters should address AI use openly, not to prohibit it, which would be unenforceable and probably unfair, but to invite transparency and explain what actually helps.
- Treat a written account as a starting point, staying alive to the difference between a party’s own words and their edited words.
- Recalibrate the credibility factors so polish is never mistaken for reliability, and clumsy writing is never mistaken for evasion.
- Authenticate digital evidence and be willing to record that an item could not be authenticated and what weight we gave it.
- Manage volume with a firm and courteous hand. Acknowledge the submission, rule on it, keep the file moving, and document the reasoning so the record shows the party was heard.
- Set expectations early about jurisdiction, the standard of proof, and what an investigation can and cannot deliver.
That last one may be the most valuable thing we can offer. Had someone told our complainant at the outset that her own words in her own voice were the most persuasive thing she had, she would never have needed the forty pages.
A closing thought
We speak often about a people-centred process, about dignity, about doing no further harm. Those commitments do not change because a party used a language model to write the complaint, and I would argue they matter more now, because the technology has created fresh ways for a party to be misjudged and fresh ways for a process to become adversarial and slow. The lesson is clear. Our tools will keep changing and so will theirs, and the only thing that reliably holds this work together is the judgment, patience, and fairness we bring to it ourselves. That is a call to action for all of us.
