Psychological Costs of AI Adoption

Sources#

Summary#

Organizational AI adoption is normally evaluated on uptake and productivity. Alami, Paja & Tiwari (University of Southern Denmark / IT University of Copenhagen, arXiv 2609.03456, 2026-09-03, case-study) ask what it costs the people enacting it, and answer with a single holistic case study — "SoftHouse", a pseudonymous 1,200-employee Danish software-services firm, one year into an organization-wide adoption — built from 21 semi-structured interviews, 5 stakeholder meetings, and 12 member-checking interviews.

They define psychological costs as "the adverse cognitive, emotional, motivational, and social experiences that practitioners incur during organizational AI adoption" and name five: uncertainty distress, accountability anxiety, cognitive load intensification, craft identity disruption, and meaning and satisfaction erosion.

The finding that carries furthest is a negative one: these costs are not resistance. "We found little evidence that practitioners opposed AI adoption itself." Where adoption stalled it was compliance constraints, uneven use-case fit, and business-as-usual delivery pressure — not opposition. The costs "coexist with acceptance, enthusiasm, and perceived productivity benefits," and they appear across the whole adoption range, from exploratory to agentic use. Any adoption model that reads slow uptake as change resistance is measuring the wrong thing.

The case, stated precisely, because the boundary conditions do the work#

Domain. Products in public health, government administration and taxation, some in operation over two decades, running in 50+ countries. Reliability, security and regulatory compliance are "contractual and legal obligations rather than aspirations," enforced by a multi-layer regime of external legislation, process certifications and internal audits. Agile since ~2005; ~50% of the workforce are software engineers.
Adoption was encouraged, never mandated. The paper is explicit and repeats it three times ("supported rather than prescribed", "encouraged but not mandated", and again in the boundary conditions). The strategy is what the authors call experiment-and-share: no prescribed workflow, projects experiment inside their own compliance envelope and the lessons are meant to circulate.
Tools. GitHub Copilot was rolled out enterprise-wide at launch (January 2025) and abandoned three months later over "quality issues" — engineers reported "code quality issues" and "overall dissatisfaction with the product." It was replaced by Claude Code and Claude Desktop, which between them account for every participant's tooling in the sample; three participants add something else (CodeScene, a local LLM, custom agents). One participant's project permitted only locally hosted models and the experiment failed on "limited computation power."
The machinery of encouragement. A dedicated AI program with an appointed director; per-project "AI ambassadors" (who "did not necessarily have prior experience with AI"); a knowledge hub; pervasive internal messaging ("every meeting is about AI... I can't remember a company briefing or anything that was not about AI"); and a five-level AI maturity model (novice → agentic) against which practitioners assess their own usage, with an organizational ambition that they climb it.
Governance was unfinished. One year in, rules existed for data classification, online access and credential delegation to agentic tools; the comprehensive internal AI policy "remained under development." Permitted use varied by customer and domain, with commercial cloud tools prohibited in some engagements.

This combination — high accountability, unfinished governance, voluntary adoption under visible measurement — is what the paper treats as the cost amplifier, and it is the part that transfers.

What the organization measured, and what it did not#

The gap between the two is the paper's practical argument.

Measured: individual position on the 1–5 AI maturity ladder (self-assessed, but visible and discussed by leadership), and an enterprise adoption-rate target — "achieving 50% adoption rate among software engineers." The target was not met at one year, which is why SoftHouse joined the research consortium in the first place.

Not measured: anything on the cost side. No productivity or efficiency measurement appears anywhere in the case description — leadership expected gains ("we really also foresee that there is productivity and efficiency gains") and tied them to a growth thesis, but the paper reports no instrument. Verification effort is not counted, not budgeted and not recognized as work. Nothing tracks identity, meaning or satisfaction. The authors' own prescription is to "decouple adoption expectations from individual measurement, and recognize verification work as legitimate effort."

The result is a measurement regime that can only report progress. Practitioners read the same instrument as a threat: "It's very offensive and uncomfortable now of course with all the measuring of people's AI maturity and that it's all over the news that this is causing people to be fired in a big way I'm sure everybody is also motivated to use AI by fear" (P1); "they defined maturity levels, and up to a level 5... So they're pushing us to evaluate us" (P19). A maturity model built as an adoption thermometer was received as a performance instrument — the single most portable organizational lesson in the paper.

The five costs, and the asymmetry between them#

The pathway figures (Figs. 4.2–4.6, viewed) encode a structural claim the prose states only in passing, and it is the most actionable thing in the paper: only three of the five costs have an organizational amplifier.

Cost	Arises from	Organizational amplifier
Uncertainty distress	Fast pace & emerging technology	Epistemic uncertainty; perceived adoption pressure
Accountability anxiety	Unknown risks of AI output	Governance ambiguity; pre-existing quality/compliance standards
Cognitive load intensification	Downstream of accountability anxiety ("translates into")	Governance ambiguity; pre-existing standards
Craft identity disruption	Role-altering capability	none
Meaning and satisfaction erosion	Role-altering capability	none

Identity and meaning have a single arrow and no amplifier box. The authors draw the consequence: "The costs rooted in the disruption itself cannot be prevented by better change management only, but they can be supported." Governance clarity, protected learning time and verification capacity are real levers on three costs and no lever at all on the two that practitioners described as irreversible.

The individual costs, in the practitioners' own register:

Uncertainty distress is not fear of AI. It is knowledge going perishable: "AI continually invalidates emerging routines before they mature," so adaptation becomes permanent rather than transitional. "I worry, how can I establish a consistent workflow in this volatility" (P7). One senior engineer's verdict on the adoption strategy itself: "experiment and tell us what you think would be no longer a valid strategy now" (P2) — experiment-and-share does not converge when the target moves faster than the sharing cycle.
Accountability anxiety is the gap between retained responsibility and the ability to discharge it. "It eases the load, but it increases the anxiety of doing something wrong" (P2); "not being able to fully vouch for the result" (P7). SoftHouse's trust-by-default culture ("we trust you until everything else is proven wrong... they trust us to be the gatekeepers", P6) is normally an asset; under governance ambiguity it "converts every practitioner into their own compliance officer" and transfers the liability judgment onto the individual. The observable failure mode is paralysis, not recklessness: "I don't know what I'm allowed to do, so I don't do anything, to not run the risk that I do something I was not allowed to do" (P11).
Cognitive load intensification is the accountability cost cashed out in hours. "I have to do more work. And it hasn't eased my cognitive load... Because now you can do more, so you feel like you need to do more... I need to look through the result line by line, I'm accountable for it" (P12).
Craft identity disruption is the loss of the activity, not of the competence. "Being a software engineer who writes code is sort of who I was, and now I'm a software engineer who prompts an AI and then reads the code" (P7). "I did not go to university to do this" (P11). An architect: "it's taking away all the fun parts... And now I'm just a prompt monkey" (P17). The authors separate this from techno-insecurity precisely: practitioners "may retain valuable expertise while losing opportunities to enact it," which is why "training or employment reassurance may leave the underlying identity disruption unresolved."
Meaning and satisfaction erosion is the authorship loss. "The final product feels like something I didn't make anymore because it's written by some machine" (P12); "the actual thing that you deliver, isn't written by me anymore. It's written by some machine that understood maybe part of what was in my head at some point" (P14). And the anticipatory version, which is the sharpest sentence in the paper on review load: "the scare for your job to become really boring, so that instead of building and creating stuff you are basically reviewing stuff, which nobody likes" (P1).

Agency displacement: the cost is proportional to how much generative work AI takes#

The paper's first named construct. Agency displacement is "the extent to which AI assumes the activities through which practitioners traditionally exercise professional judgment, expertise, and ownership" — and cost prevalence tracks it, not adoption level and not role seniority.

The role-based magnitude analysis (Table 7, counting participants with ≥1 first-cycle code per cost, reconciled against pdftotext -layout):

Role	Accountability anxiety	Craft identity	Meaning erosion	Cognitive load	Uncertainty distress
Software Engineers (n=10)	10/10	10/10	10/10	7/10	10/10
Software Architecture (n=3)	0/3	1/3	1/3	1/3	0/3
Testing (n=3)	1/3	0/3	1/3	0/3	3/3
Senior Management (n=5)	2/5	0/5	0/5	0/5	3/5

The engineer row is saturated: every software engineer in the sample voiced four of the five costs. The other rows are not a softer version of it — they are different shapes.

Architects came out ahead. AI extended their generative reach instead of taking it: "I suddenly had a developer in my back pocket" (P17, of Claude Code as a prototyping partner). A colleague's less charitable reading of architect enthusiasm — "they don't have a developer saying no to them" (P12) — is the same fact from the other side: what architects gained is exactly the execution capacity that used to require negotiating with an engineer.
Testers are the counterexample that keeps agency displacement honest. AI complements evaluative work, so craft identity disruption is 0/3 and one tester reports unchanged enjoyment ("it complements my effort... I really like working with it"). Yet uncertainty distress is 3/3 — the highest fraction in the table outside the engineer row — because nothing in the experiment-and-share stream applied to their work: "the ones [experiments] that were shared across a lot of us weren't really relevant for my role." A role can be untouched by displacement and still bear the full organizational cost. The tester's standing concern is separate again and very direct: "a tester is easier to fire than an architect" (P1), which is why she declined to experiment with deeper integrations at all.
Senior management carries accountability anxiety and uncertainty distress with zero identity or meaning cost — they own the initiative and must show it working.

The authors' own caveat, which the numbers invite readers to forget: these fractions come from 3-person and 5-person cells, saturation was assessed at sample level and not within roles, and the analysis is explicitly "descriptive... exploratory rather than conclusive." They support no statistical comparison across roles. The direction is the claim; the ratios are not.

Practical consequence the paper draws: "a single AI adoption strategy is unlikely to be equally effective across software engineering roles" — engineer-facing adoption needs preserved generative work, architect- and tester-facing adoption needs decision augmentation.

The construct predicts an attitude it was not built to predict (2026-09-23). Agency displacement says the cost lands in proportion to how much of a role's generative work AI takes. If practitioners know that, resistance should be strongest exactly where displacement would be deepest — and in a population with no overlap with this one, measured by a method with nothing in common with this one, it is. Angelini & Lyrvall's latent class analysis of 3,785 PhD students (Normative boundaries of AI in scientific work: Evidence from PhD researchers, arXiv 2608.25678, empirical) finds comfort with AI split along a task boundary rather than an accept/reject one: 51.5% comfortable with AI summarising literature against ~30% for writing a paper, analysing data or designing experiments — the three tasks that carry authorship and scientific judgment, i.e. the generative half of the work. Their dominant 44% profile is never enthusiastic about anything and still keeps that gradient. The papers reach the same object from opposite directions: this one measures what displacement costs people already living it, that one measures where people who can still choose refuse to let it happen. Both name the same residual — what §5.2 there calls the tasks "through which researchers learn to formulate arguments, recognise limitations, evaluate evidence, and make independent decisions."

The verification tax#

The second named construct, and the one that connects this page to the rest of the wiki's review-load evidence. The verification tax is "the additional cognitive and engineering effort required for practitioners to remain accountable for their work when AI-assisted." It is explicitly not a sixth cost but "the direct synthesis of the accountability anxiety and cognitive load intensification" — the figures draw it as a single "translates into" edge from one to the other.

The mechanism is stated by a participant more cleanly than by the literature: "if we weren't responsible for the code it produced, it would be a lot faster" (P19). The tax is levied by retained accountability, not by output volume. Nothing about the AI's throughput sets its rate; the rate is set by how much understanding a practitioner must reconstruct before they are willing to sign.

The authors' sharpest observation about why this is expensive rather than merely tedious:

"Software engineers do not verify code because they finished writing it; they verify code while they are writing it.... By externalizing code generation, AI compresses this process into predominantly post hoc verification, requiring practitioners to evaluate solutions without having participated fully in the reasoning that produced them."

Verification was never a separable phase whose cost could be measured on its own. Delegating authorship creates it as a phase, then bills for the understanding that used to accumulate for free. This is the human-side mechanism for the review burden that Verification as the New Bottleneck and Acceleration Whiplash measure from telemetry, and it is why the load does not fall when the model gets better at writing code — it falls only when the practitioner needs less reconstructed understanding to stand behind the output.

The cost is also tied to specific practices rather than floating free. Line-by-line reading before acceptance (P12); analysis of every generated solution before adoption (P2); and, in peer review, heightened scrutiny applied specifically to AI-authored contributions (P16) — AI provenance changes what a reviewer does, inside a code-review process that was designed on the assumption that a colleague wrote the diff.

Two implications the authors put in industry terms: "the productivity gains of AI cannot be measured solely by code generation speed," and organizations "must explicitly account for the increased cognitive load of code verification in capacity planning and sprint estimates."

Manage, mitigate, absorb — and the one cost nothing reaches#

Responses are organized by function, not behaviour, around the three resources being protected (Table 8, reconciled against the prose):

Resource	Function	Practices
Agency (being in control)	Manage	Staying in the loop; prior codebase knowledge as an oversight prerequisite; complexity-based delegation triage; plan and diff review; baseline testing before accepting changes; deliberate deceleration; line-by-line code verification
Agency	Mitigate	Self-imposed limits pending explicit permission; discontinuing unsanctioned experiments; stripping data before cloud tool use; partitioning work into safe and grey zones; declining write-capable integrations ("MCP integration is so far a no-go. I won't do that, because I won't get fired", P9)
Competence (being capable)	Mitigate	Deliberate periodic manual-coding ritual ("if you don't use it, you lose it", P7); self-then-AI sequencing (own draft before consulting AI, to protect learning); delegation moderation — boilerplate to AI, business-logic-heavy work retained
Professional identity	Mitigate	Role redefinition toward orchestration; relocating worth from writing code to "delivering software, not delivering code" (P10); positioning the human as guardian of architectural integrity; reframing SE as problem solving
Professional identity	Absorb	Enduring the loss of enjoyment; displacing enjoyment outside paid work; continuing without resolution; capitulation

Absorption is the paper's genuine theoretical addition — a response category that is neither coping nor resistance, in which the loss is appraised as irreversible and simply carried. Asked how he handled the increased mental burden, P7 answered: "I don't. I just keep going." P19 on losing the enjoyment of coding: "not a nice feeling."

The asymmetry across resources is the finding. Agency is restorable by changing how work is done; competence by changing how learning happens; identity by neither. Manage and mitigate reach agency and competence. Identity gets mitigation (reframing) or absorption, and nothing else — which is the same conclusion the pathway figures reach from the other direction.

Note that the manage column is also the verification tax's invoice. Every agency-preserving practice is verification labour, so an organization that succeeds in getting practitioners to "stay in the loop" has by construction increased the cognitive load it is not measuring.

How much weight this deserves#

case-study, correctly tiered, and the design supports explanatory transfer rather than magnitude claims.

Strengths worth naming. Access during rather than after the transition; five stakeholder meetings reconstructing the adoption history independently of the interviews; 12 of 21 participants did member checking against a 15-page findings report and did not require changes to the interpretations or the model; participants were anonymous to the employer except the AI program director; the role-magnitude analysis counts participants rather than coded segments, so a voluble participant does not outweigh a terse one; and the authors write four pages of boundary conditions qualifying their own transfer claims.

Limits, most of them the authors'.

One company, one year, one snapshot. Cross-sectional: it cannot say whether these costs decay as governance matures or deepen as delegation does, and the authors explicitly decline to imply either. Whether the absorbers burn out or the feared skill atrophy materializes is named as unknown over a 3–5 year horizon.
The sample is old-timers. Tenure median 9.5 years (range 1–23); industry experience median 20 (range 2–40); 19 of 21 have a decade or more. The authors concede this "may have foregrounded costs associated with disrupted expertise and craft identity." A sample of practitioners whose identities formed after generative AI would plausibly report a different set — the one junior engineer (P20, 2 years' experience) is a sample of one.
Role cells are tiny (3, 3, 5) and skewed toward engineers, who are also the subjects of the most confident claims.
Prevalence is a coding artifact as well as a phenomenon. "At least one first-cycle code" is a low bar; 10/10 means every engineer said something codable, not that every engineer was distressed.
Regulated-domain amplification. SoftHouse's compliance culture is an amplifier by the paper's own model. The authors argue regulation is not a prerequisite — comparable tension "may arise wherever practitioners remain responsible for software they cannot adequately understand, validate, or justify" — but the magnitudes are this firm's.
Voluntary adoption is a boundary condition on the responses, not the costs. Manual-coding rituals, bounded delegation and declined integrations all require discretion over how work is done. Where AI use is prescribed, the paper says plainly that it "does not establish how practitioners respond."
No gender analysis (5 women, 16 men), declined for representation reasons.
COI is light and disclosed: recruited through the AI4SE Denmark consortium, funded by Innovation Fund Denmark, with the first author both interviewer and relationship-holder for the company and the AI program director both gatekeeper and participant (P21) — a recruitment channel that selects for practitioners willing to be seen talking to researchers about AI.

The population counterpart, and the two costs it cannot see (McKinsey, August 2026)#

The state of AI in 2026 (McKinsey / QuantumBlack, 2026-08-25, empirical but self-reported, 1,719 respondents in 97 nations) is the largest instrument in the corpus that asks practitioners how AI use affects them personally, and reading it against this 21-person case study is instructive in both directions.

Where it corroborates. Negative effects are real, minority, and concentrated below the executive line (Exhibit 5): 47% of midlevel managers and individual contributors report at least one negative effect, against 31% of executives and senior managers. Item by item, midlevel manager / individual contributor / C-level / executive-senior-manager: "hurts my ability to think critically" 20 / 16 / 9 / 8; "increases anxiety about my career prospects" 19 / 18 / 6 / 10; "creates pressure to take on more work" 18 / 10 / 12 / 11; "increases my mental fatigue" 14 / 7 / 7 / 7. Positive items are flat across levels (productivity 81 / 76 / 80 / 81), so the gradient is entirely on the cost side — which is the population-scale form of this study's central asymmetry between who reports benefits and who carries costs. The skill-atrophy fear this cohort named unprompted has a population correlate in the critical-thinking row, and the fear-driven adoption it describes has one in the career-anxiety row (13% overall).

Where it is blind, which is the more useful finding. The survey's instrument has thirteen impact items — seven positive, six negative — and none of them is craft identity, meaning, agency or accountability. The two costs this study's participants appraised as irreversible, and the one the study found no organizational amplifier for, are invisible to the largest self-report instrument the field runs annually. A firm reading only the McKinsey series would conclude that AI's personal costs are pressure, anxiety, fatigue and output volume — all of which are manageable-category items in this study's taxonomy — and would never learn that the losses practitioners called permanent exist. That is a measurement gap, not a disagreement: the interview design found the costs precisely because it did not start from a fixed item list.

Connections#

Layered Supervision — the organizational half of the same re-scoping, reported a month earlier and from five Swiss teams instead of one Danish firm. Where this page finds that retained accountability converts into line-by-line verification labour, that source finds practitioners abandoning line-by-line reading as impossible and substituting operational explainability — keeping the system diagnosable and reconstructable rather than understood. The two are the same pressure resolved in opposite directions, and the split may be the compliance regime: a regulated shop cannot trade comprehension for diagnosability. Both find the burden falling on engineers and not architects
Community Smells Under AI Adoption — the same question (what does AI adoption do to the human system it lands in?) asked with the opposite instrument and answered in the opposite direction. That page's PLS-SEM over 152 professionals finds fewer socio-technical anti-patterns under AI adoption; this finds costs everywhere. The two are compatible and the reason is structural: community smells are team-level properties measured by interaction frequency and communication quality, and none of the five costs here is a team-level property. An engineer can consult colleagues more often, communicate better, and still be absorbing craft identity loss in silence — which is exactly what P12's member-checking line describes ("Some of your findings have been talked about in the corridors and during lunches, but we never openly confronted them"). Where the two instruments do touch, this source supplies the mechanism for the knowledge-sharing friction the smells model cannot see
Verification as the New Bottleneck — the hub this page supplies the human cost for. The verification tax is that bottleneck's bill, and it prices differently than the telemetry does: not in review hours but in the accountability that makes those hours unskippable. P19's "if we weren't responsible for the code it produced, it would be a lot faster" is the cleanest statement in the corpus of why the bottleneck is where it is
Acceleration Whiplash — the telemetry and the interviews describe one phenomenon from two ends. Faros counts review time up 5x and incidents per PR up 243%; this counts the practitioner reading the diff line by line because they are accountable for it. This source adds the part the telemetry cannot see — that the load is borne differently by role, and that the engineers bearing most of it are also the ones losing the work they valued
Outsource Your Thinking, Not Your Understanding — the workplace instance of that page's thesis, arrived at independently. "Verification requires them to reconstruct sufficient understanding of delegated work to assess and endorse it" is the same non-delegable residue, and this source prices it: the reconstruction is the cost, and it is expensive precisely because authorship used to supply the understanding for free
The Tragedy of the Cognitive Commons — that page's depletion mechanism, observed as first-person anticipation rather than measured outcome. Practitioners name the mechanism unprompted ("if we lose access to these things... will I still be good at it?", P7; "if you use it too much, you will forget some of the skills", P20) and counter it with deliberate manual-coding rituals — individual-level conservation of exactly the resource that page models as a profession-level commons. It cuts both ways: the practitioners are defending the commons voluntarily, which is a better outcome than the model predicts, and they are doing it on their own initiative with no organizational support, which is worse
Role Averaging, Not Role Elimination — "your role is the average of what you spend your time on," experienced from inside as loss. This is that thesis's cost side: P7's before-and-after ("a software engineer who writes code" → "a software engineer who prompts an AI and then reads the code") is an averaging event described by the person being averaged. It also supplies the organizational instrument that page's ladder question is about — a five-level AI maturity model applied to individuals
Pilot-to-Production Gap — its Consideration 05 (converting operators into overseers) has an oversight workforce here, and the conversion is neither free nor stable. Overseers who remain accountable for output they did not author pay the verification tax, and the ones whose craft was the displaced execution work also absorb identity and meaning losses that no amount of oversight training addresses
AI Brain Fry — the measured consequence of the strain described here. Kropp et al. find that workers fatigued by AI oversight make 11% more minor and 39% more major errors; this supplies the reason the oversight is not negotiable down, which is retained accountability rather than output volume. The pairing is uncomfortable and worth holding: every agency-preserving practice in the manage column above is oversight volume, so an organization that successfully keeps humans meaningfully in the loop has bought the error rate that page prices
Telemetry vs. Survey Measurement — a third instrument class with a distinct visibility profile. Interviews plus member checking see appraisal, identity, meaning and absorbed strain, none of which telemetry or a Likert survey can reach; they cannot see magnitude, trend, or anything about practitioners who declined to volunteer
AI Adoption in Scientific Work — agency displacement arriving from the other side of the decision. Where this page measures the cost to people whose generative work AI has already taken, that page measures 3,785 PhD students drawing the line before it is taken, around the same tasks: literature work accepted (51.5% comfortable with AI summarising), writing, analysis and experiment design resisted (~30%)
Engineer PM Convergence — the convergence's quieter direction. "I suddenly had a developer in my back pocket" is an architect absorbing engineering execution, and "they don't have a developer saying no to them" identifies what actually changed: not the architect's skills but the removal of the human gate that used to bound their scope

Open Questions#

Absorption is a response category defined by the loss being appraised as irreversible, and the paper cannot say what it turns into. Do practitioners who reported absorbing craft-identity and meaning losses at one year show elevated attrition, burnout or disengagement at three — or does absorption stabilize into a new baseline once the comparison to pre-AI work stops being available? The design that settles it is the authors' own: track this cohort, not a new one.
The pathway figures give craft identity disruption and meaning erosion no organizational amplifier, which is a strong claim derived from one firm's coding — it implies no adoption strategy can touch the two costs practitioners called irreversible. Does a firm that deliberately preserved generative work for engineers (protected non-delegated tasks, bounded AI scope by role) show lower identity and meaning costs at comparable adoption depth, or do the two costs track raw exposure regardless of work design? #oq/source The intervention gets a second, independent prescription — and still nobody has run it (2026-09-22): When Does AI Augment Work? A Workflow-Level Framework for Human-Agent Collaboration (CIVIC-AI workshop whitepaper, practitioner-opinion, no measurement) makes job purpose one of six conditions for genuine augmentation and states this bullet's concern as that condition's failure mode almost verbatim — "AI does desirable tasks, while humans are relegated to undesirable ones" — while its worked example prescribes exactly the protected-generative-work arm the bullet asks for: juniors perform the initial protocol design and coding themselves rather than reviewing an agent's, on anti-anchoring and anti-vacuous-verification grounds. The useful addition is the argument for running the comparison, which is not a morale argument: the framework puts learning and purpose upstream of meaningful human control, making protected generative work a prerequisite for the firm's own oversight integrity rather than an equity concession. No firm, no measurement, no matched adoption depth — the bullet stands as posed.
Every magnitude in the role analysis rests on a sample whose industry experience has a median of 20 years. Do practitioners whose professional identity formed after generative AI report craft identity disruption and meaning erosion at all, or does the cost profile collapse to the accountability/verification half? A replication stratified by career-entry cohort would separate a transition cost from a permanent one, and the paper's 10/10 engineer rows cannot.

Sources#

The state of AI in 2026: On the road to ROI — Dan Tinkoff, Lieven Van der Veken & Michael Chui with Tara Balakrishnan, The state of AI in 2026: On the road to ROI (McKinsey / QuantumBlack, 2026-08-25, empirical, self-reported; online survey, 1,719 participants in 97 nations, fielded May 4 - June 8 2026, GDP-weighted). Cited here for Exhibit 5 (perceived impacts of AI on the respondent's own work, by organizational level) and the 13% overall career-anxiety figure. Its thirteen-item impact battery contains no identity, meaning, agency or accountability item, which is why it can only corroborate the manageable half of this page's taxonomy. COI: McKinsey sells AI transformation consulting; the panel is its own.
The Psychological Costs of Artificial Intelligence Adoption in Software Engineering — Adam Alami, Elda Paja & Abhishek Tiwari (University of Southern Denmark; IT University of Copenhagen), arXiv 2609.03456, 2026-09-03, case-study, 43 pages. §3.2 case description and the adoption context (Copilot abandoned at three months, Claude Code/Claude Desktop, AI ambassadors, the 1–5 maturity model, the unmet 50%-adoption target, governance still under development); §3.3 and Table 2 the participant sample; §4.1–4.5 the five costs with their pathway figures; §4.6 and Table 7 the role-based magnitude analysis; §4.7 and Table 8 the manage/mitigate/absorb response taxonomy; §5.3 boundary conditions; §5.4 agency displacement and the verification tax; §7 limitations. Figures 4.1 (integrated model), 4.3, 4.4, 4.5 and 4.6 viewed under the inline-image rule — the amplifier asymmetry in the table above comes from the figures, not the prose. Parse notes. PDF-derived (docling 2.126 / MLX). The ingest table-collapse (6 cells) and table-weld (5 cells) warnings both trace to Table 2 and are false positives on inspection — comma-separated tool lists and two-line role labels — with one real weld: P18's role reads "Architect Lead SE" in the parse and "Lead SE" in the reference parse. Table 4's grouped sub-rows are collapsed in the docling grid (three quotes in one cell) but recover in order against pdftotext -layout. Tables 2, 4, 7 and 8 were each reconciled against pdftotext -layout before anything on this page was quoted from them; no unreconciled row is cited. Running-header fragments ("The Maersk Mc-Kinney Moller Institute") are mis-tagged as ## headings through §4 in the raw and are not section boundaries. A source-side inconsistency, not a parse artifact: the tester's experiment-and-share quote is attributed to P10 in §3.2 and to P11 in §4.1, in a sentence that names P9 — confirmed in the reference parse. It is quoted here without a participant ID
Normative boundaries of AI in scientific work: Evidence from PhD researchers — Angelini & Lyrvall, Normative boundaries of AI in scientific work: Evidence from PhD researchers, arXiv 2608.25678, 2026-08-26, 15pp, empirical. Cited here only for the convergence with agency displacement: the Table 1 marginals, the four-class solution (§4) and the skill-formation argument in §5.2. It is a self-selected international survey of PhD students in STEM and medical/health sciences measuring comfort, not behaviour and not cost; full treatment and caveats on AI Adoption in Scientific Work