Employ­ers need to be aware of legal risks around AI prompts.

Even where no priv­ileged com­mu­nic­a­tion is involved, AI prompts and out­puts may become rel­ev­ant and pro­du­cible in lit­ig­a­tion. — Howard Levitt and Jef­frey Buchan

Arti­fi­cial intel­li­gence has already divided the busi­ness world into two camps: those using it and those, well, rap­idly fall­ing behind.

For employ­ers, exec­ut­ives and pro­fes­sion­als, refus­ing to use AI is no longer prudence. It is com­mer­cial self-sab­ot­age. AI can accel­er­ate research, improve draft­ing, expose weak­nesses in an argu­ment and reduce the time required to make dif­fi­cult decisions.

But its ease of use has cre­ated a dan­ger­ous illu­sion: that typ­ing into an AI plat­form is the digital equi­val­ent of think­ing out loud.

It is not.

The prompts an employer enters, the doc­u­ments it uploads and the answers it receives may become evid­ence in lit­ig­a­tion. Worse, insert­ing legal advice or con­fid­en­tial com­mu­nic­a­tions into an open AI plat­form may jeop­ard­ize soli­citor-cli­ent priv­ilege — the priv­ilege courts afford to com­mu­nic­a­tions between law­yer and cli­ent to retain its con­fid­en­ti­al­ity and avoid dis­clos­ure to the oppos­ing side and to the court.

The seem­ingly private exchange tak­ing place on a laptop today could become the most dam­aging exhibit in a courtroom sev­eral years from now.

Two risks deserve par­tic­u­lar atten­tion: the inad­vert­ent waiver of soli­citor-cli­ent priv­ilege and the pro­duc­tion of AI prompts and out­puts dur­ing lit­ig­a­tion.

In employ­ment cases, where an employer’s motive and decision-mak­ing pro­cess are often cent­ral, those risks are espe­cially acute.

Soli­citor-cli­ent priv­ilege is among the most jeal­ously pro­tec­ted prin­ciples in Cana­dian law. It per­mits cli­ents to speak can­didly with their law­yers, secure in the know­ledge that their com­mu­nic­a­tions will remain con­fid­en­tial.

But priv­ilege depends upon con­fid­en­ti­al­ity. When pro­tec­ted inform­a­tion is vol­un­tar­ily dis­closed to an out­side party, the priv­ilege is waived.

The prob­lems often arise inno­cently. A cli­ent receives advice from coun­sel and wants a second opin­ion. Or a cli­ent wants an AI tool to sim­plify a legal memor­andum, chal­lenge the law­yer’s ana­lysis or sug­gest a bet­ter set­tle­ment strategy.

To obtain a use­ful response, the cli­ent sup­plies the facts. They may quote from their law­yer’s email, upload cor­res­pond­ence, describe set­tle­ment dis­cus­sions or identify per­ceived weak­nesses in the case.

At that point, the user may have dis­closed pre­cisely the inform­a­tion the priv­ilege was inten­ded to pro­tect.

Cana­dian courts have not yet provided guid­ance on whether using a AI in this fash­ion amounts to waiver. But a recent U.S. decision offers a warn­ing.

In that case, a lit­ig­ant entered inform­a­tion received from his law­yers into an AI tool, used it to gen­er­ate doc­u­ments and then provided those doc­u­ments to coun­sel for the lit­ig­a­tion. The court con­cluded that dis­clos­ure to an open AI plat­form amoun­ted to dis­clos­ure to a third party and that priv­ilege had there­fore been lost.

Amer­ican priv­ilege law is not identical to Cana­dian law, and a U.S. decision does not bind a Cana­dian court. But that is scant solace. No employer or exec­ut­ive should aspire to become the Cana­dian test case.

The con­sequences can be severe. Advice about legal expos­ure, lit­ig­a­tion strategy, set­tle­ment ranges and the strengths or weak­nesses of a case may become dis­clos­able to an oppos­ing party.

A few care­less minutes with an AI plat­form might com­prom­ise months or years of pro­tec­ted com­mu­nic­a­tions.

For employ­ers, the danger extends well bey­ond priv­ilege.

Even where no priv­ileged com­mu­nic­a­tion is involved, AI prompts and out­puts may become rel­ev­ant and pro­du­cible in lit­ig­a­tion.

Employ­ers are increas­ingly turn­ing to AI for assist­ance with ter­min­a­tions, per­form­ance man­age­ment, har­ass­ment com­plaints, dis­ab­il­ity accom­mod­a­tion, restruc­tur­ing and employee dis­cip­line.

The appeal is obvi­ous. Why pay a law­yer when a plat­form can pro­duce an imme­di­ate answer at little or no cost?

Why? Because the “free” answer may later become an extremely expens­ive exhibit.

Employ­ment cases fre­quently turn on intent. Courts and tribunals exam­ine what the employer knew, when it knew it and why it acted as it did. They scru­tin­ize internal emails, draft memor­anda, meet­ing notes and com­mu­nic­a­tions among decision-makers.

An AI prompt may offer a very can­did record.

Con­sider a com­pany con­tem­plat­ing the dis­missal of a long-ser­vice employee. A man­ager might ask: “How do we ter­min­ate this employee while pay­ing the least pos­sible sev­er­ance?” Or: “How can we avoid a human rights com­plaint?” Or: “Can we elim­in­ate an employee’s pos­i­tion while she is on dis­ab­il­ity leave?”

The man­ager may regard those ques­tions as pre­lim­in­ary research. Oppos­ing coun­sel may char­ac­ter­ize them very dif­fer­ently. And a court may find the ques­tion damning.

A prompt framed care­lessly could be used as evid­ence of bad faith, dis­crim­in­at­ory intent, retali­ation or a delib­er­ate attempt to evade stat­utory oblig­a­tions.

Even if that char­ac­ter­iz­a­tion is unfair, the employer will have sur­rendered enorm­ous nego­ti­at­ing lever­age. Instead of lit­ig­at­ing the mer­its of the dis­missal, it may be tasked with explain­ing an ugly-look­ing prompt cas­u­ally draf­ted months earlier by someone who never expec­ted any­one else to read it.

The AI out­put can cre­ate fur­ther prob­lems. An employer may rely on an answer that is leg­ally wrong, based on advice from another jur­is­dic­tion or obli­vi­ous to facts that would have changed the ana­lysis.

AI writes with con­fid­ence even when it is mis­taken. That com­bin­a­tion is dan­ger­ous in employ­ment law, where the res­ult fre­quently depends on the word­ing of a con­tract, the employee’s jur­is­dic­tion, the employer’s prior con­duct, the most recent case law and the pre­cise chro­no­logy of events.

None of this means that busi­nesses should stop using AI. That would be both unreal­istic and com­mer­cially fool­ish.

It means they must gov­ern its use with the same ser­i­ous­ness they apply to email, doc­u­ment reten­tion, cyber­se­cur­ity and legal com­mu­nic­a­tions.

Employ­ers should adopt clear rules pro­hib­it­ing employ­ees from enter­ing priv­ileged advice, con­fid­en­tial legal cor­res­pond­ence, per­sonal employee inform­a­tion or sens­it­ive busi­ness records into unap­proved AI plat­forms.

They should also require employ­ees to assume that every prompt may one day be reviewed by a judge, reg­u­lator, invest­ig­ator or oppos­ing law­yer.

Before press­ing “enter,” the user should ask:

❚ Would I be com­fort­able hav­ing this prompt read aloud in court?

❚ Would I be pre­pared to explain why I asked the ques­tion in this way?

❚ Does this prompt dis­close con­fid­en­tial, priv­ileged or per­sonal inform­a­tion?

❚ Am I using AI to assist judg­ment, or to replace the pro­fes­sional advice the situ­ation actu­ally requires?

AI can be an extraordin­ar­ily use­ful tool. But it is not a con­fes­sional, and it is not a priv­ileged con­ver­sa­tion simply because no one else is in the room.

The most dam­aging evid­ence in the next gen­er­a­tion of employ­ment cases may not be found in an email, text mes­sage or board memor­andum.

It may be found in a prompt someone wrongly assumed no one would ever see.

Howard Levitt is senior part­ner of Levitt LLP, employ­ment and labour law­yers with offices in Ontario, Alberta and Brit­ish Columbia. He prac­tises employ­ment law in all provinces and is the author of six books, includ­ing The Law of Dis­missal in Canada. Jef­frey Buchan is an asso­ciate at Levitt LLP.