Does expertise make AI feel more threatening, not less?

← Back to Blog

Identity threats are “experiences appraised as indicating potential harm to the value, meanings, or enactment of an identity.”

Petriglieri (2011), Academy of Management Review

In a 2,257-person sample I worked on, AI threat perceptions and depth of use varied significantly by job level rather than distributing evenly across the workforce (Reich et al., 2026) (Emerging). And Shonhe and Min (2025) found the split running along expertise: as implementation recedes in time, professional identity threat intensifies among expert practitioners while use intention falls among novices (Emerging).

Both point the same direction. The people carrying the most expertise are the ones whose use stalls, and they stall for reasons unrelated to not knowing how the software works. Picture the shape of it in a legal department. Associates route first-pass contract review through a new analysis tool several times a day. The general counsel opens it twice a month, for questions she already knows the answer to. Not because the tool is bad, but because reading the contract is what her expertise consists of, and the tool proposes to take over the reading. The activation report records both behaviors as adoption.

My argument here: professional identity threat concentrates among senior experts whose autonomy and judgment AI encroaches on. Because that threat is an appraisal of role scope and decision rights rather than a skills deficit, it caps depth of use long after adoption has formally begun. That makes it a collaborative intelligence problem, not a sentiment to manage through better change communications.

The Research

Identity threat has a long theoretical pedigree in organizational psychology, and it has never been about feelings in the loose sense. Petriglieri (2011) frames it as an appraisal process: people assess whether an experience devalues, discontinues, or delegitimizes an identity they hold, then act to protect either the identity or its meaning (Established). The appraisal is about role legitimacy. That matters for how we read AI adoption data, because an appraisal responds to structure, and structure is something organizations control.

Recent workplace research gives the construct empirical teeth and names the mechanism. Ashraf, Min, and Ashraf (2026), in a three-wave survey of 507 employees, link perceived loss of skill and autonomy to identity threat and to downstream disengagement, including cyberloafing (Emerging). That is the inside view of what the two opening findings observe from outside: what varies by seniority is not comfort with software but how much of a person's discretion the system proposes to absorb.

Read together, this is a collaborative intelligence question. Collaborative Intelligence is the capability that emerges when human judgment and machine capability combine and exceed what either produces alone, and it only emerges where a person actually routes a judgment call through the system. Confirmation traffic does not produce it. If threat tracks seniority and autonomy rather than technical familiarity, the lever is not the training curriculum. It is the design of the role the expert returns to once the tool is in place: which judgment calls remain hers, and which ones she now audits. Kellogg, Valentine, and Christin (2020) document algorithmic systems as contested terrain over control, shifting decision rights away from workers through mechanisms that rarely appear in any adoption plan (Established). In most rollouts I see, nobody has drawn that line on purpose, so the tool draws it by default.

What This Means in Practice

Most adoption dashboards I see measure the wrong thing with impressive precision. Weekly active users, licenses activated, prompts per seat. Those metrics answer whether people opened the tool. They do not answer whether anyone routed a judgment call through it. A radiology group can post near-total activation while every attending uses the system to confirm a read they have already made, which is adoption in the reporting sense and nothing in the clinical sense.

The second failure is sequencing. Organizations roll out, hit friction among senior staff, then commission a change-communications workstream to address it. By then the expert has made the appraisal, and the appraisal was correct: her scope did shrink and nobody told her what replaced it. Raisch and Krakowski (2021) argue that automation and augmentation are a deliberate choice management must make rather than an outcome it discovers (Established). That choice belongs before deployment. Decide which judgment calls the senior underwriter still owns outright, which ones the model advises on, and which ones the model decides while she audits the exceptions. Write it down. Then ship the tool.

Third, reassurance calibrates nothing. Telling a twenty-year expert that AI will not replace her is a claim she cannot verify and has every reason to discount. What she can verify is whether the system shows its reasoning and whether her override changes anything downstream. Shonhe and Min (2025) find explainability, functioning as a visible collaborator, central to reducing professional identity threat (Emerging). That is Trust Architecture doing work a town hall cannot.

There is a fourth failure, and it happens before the tool exists. Task inventories, workflow interviews, AI opportunity assessments: all of them ask an expert to describe her own work to a program that may automate parts of it, and that request is itself an appraisal of whether the role stays legitimate and continuous (Petriglieri, 2011) (Established). So the answers come back shaped by the appraisal. My read (Speculative): under threat, the judgment content of a role inflates and the routine content deflates, and nobody has to lie for it to happen. Each step gets described in its hardest version, the fifteen minutes of pattern recognition rather than the two hours of retrieval that set them up. Which means the automation map you build from those interviews is least accurate exactly where expertise runs deepest, and that is the part you were relying on it to get right.

Three Things to Take Away

Audit depth of use by seniority and role, not binary uptake

Surface metrics hide stalled use exactly where expertise concentrates. Reich et al. (2026) found threat and depth-of-use patterns differing significantly by job level across 2,257 respondents (Emerging), which means a single organization-wide adoption number averages across groups with opposite dynamics. Segment frequency and duration by role level before concluding that a rollout worked.

Redesign decision rights before rollout, not after resistance

Treating senior resistance as a training gap misreads what is being contested. Kellogg, Valentine, and Christin (2020) show algorithmic systems redistributing control whether or not anyone planned for it (Established), and Raisch and Krakowski (2021) place the automate-or-augment decision with management (Established). Name the judgment calls your experts retain, in writing, before the tool arrives.

Build calibration loops, not reassurance campaigns

Experts need to see how their judgment is weighted against the model's, and to change that weighting when it is wrong. Ashraf et al. (2026) recommend preserving human discretion and override authority as a practical mitigation for identity threat (Emerging). An override that disappears into a log is not discretion; make it visible and make it consequential.

My Two Cents

We have been quietly relieved to call this a sentiment problem, because sentiment is cheap to address and decision rights are not. Redrawing what a senior expert owns means someone has to decide, on the record, that a category of judgment now belongs to a model. That is a governance decision with liability attached. Running a workshop about growth mindset is far more comfortable.

Here is the part I hold more loosely (Speculative): identity threat among senior experts is often a better signal than the adoption metrics sitting next to it. Low usage from your most experienced people is usually accurate reporting on an unresolved question. They are reading the structural change faster than the rollout plan acknowledges it. Organizations that treat that signal as data, and then redesign the role, will get depth. Organizations that treat it as an attitude will get 100% activation and nothing underneath.

Read to Learn More

Ashraf, Min, and Ashraf (2026), in the European Journal of Investigation in Health, Psychology and Education, model how AI-driven identity threats translate into disengagement and test AI-inclusive identity as a moderator. The three-wave design makes it stronger evidence than most work in this area.

Parsons (2026), writing for Employer Branding News, makes the employer-side case that identity, not redundancy, is the live AI threat at work. Useful as a read on how fast this framing is reaching practitioners, though it is commentary rather than evidence.

References