Resources / 3, 5, 6: Organizational Psychology / AI at Work

Maya at an office desk beside the bunny at a tiny desk wearing a necktie

AI at Work

3, 5, 6: Organizational Psychology

Study guide by Anders Chan, PsyD · Updated

Why AI at Work Is I-O Material

When an organization adds an AI tool, the hard problems are rarely the software. They are human problems I-O psychology has studied for decades: trust in machines, responsibility, job change, skill building, fairness, and monitoring. In November 2026 APA published Deploying AI Into Workplaces: What Leaders Need to Know, which organizes these problems into seven human objectives. The EPPP tests the classic concepts underneath them, so this lesson teaches the concepts and uses the APA guidance as the current example.

Think of the AI tool as new equipment on the factory floor. The engineering is the vendor's job. Whether people use it well is an I-O question.

What the APA Guidance Is

  • It is advisory guidance, not a compliance standard, not an APA professional standard, and not a rating of any product.
  • Personnel selection and other formal HR decisions are out of scope. SIOP (APA Division 14) covers AI-based hiring assessments.
  • Each objective rates its evidence: the human behavior is "well established" from decades of automation research, while effects specific to current AI tools are "emerging."

The Seven Objectives

ObjectiveCore ideaKey concepts
1. StakesMatch safeguards to the stakesError detection, common-mode failure
2. Trust and relianceTrust should match what the tool can doAutomation bias, algorithm aversion
3. AccountabilityOversight must be real, not a formalityDiffusion of responsibility
4. Job designRedesign the work, not just the task listJob demands-resources model
5. ExpertiseProtect how people build deep skillDesirable difficulties, deskilling
6. Fair processIntroduce AI through a fair processProcedural justice, psychological safety
7. MonitoringSet limits on monitoringElectronic performance monitoring

Stakes: Match Safeguards to the Stakes

A tool that summarizes meetings needs lighter checks than one that makes consequential decisions. People are poor at forecasting rare, severe, or slow-building outcomes, so for high-stakes uses the guidance says to write the consequences down rather than leave them to intuition. Beyond "How likely?" and "How bad?", it adds four questions: Who bears the consequences? Can the error be undone? How quickly would anyone know? How far could it spread?

The last question matters because one AI tool used the same way across many cases can repeat the same error in all of them. This is a common-mode failure: one cause, many simultaneous failures. A successful low-stakes pilot does not show a tool is ready for high-stakes use.

Trust and Reliance

Parasuraman and Riley (1997) named the classic ways people relate to automation:

  • Misuse: over-reliance. People stop monitoring and miss the tool's errors.
  • Disuse: under-reliance. People ignore a tool that would have caught errors, or waste time re-checking reliable work.
  • Abuse: automating without regard for the consequences for human performance.

Automation bias is the tendency to over-rely on automated aids. It produces errors of commission (following wrong advice) and omission (missing a problem because the aid did not flag it). The opposite pattern is algorithm aversion: Dietvorst, Simmons, and Massey (2015) found that people avoid an algorithm after seeing it err, even when it outperforms humans. The goal is calibrated trust (Lee and See, 2004): reliance that matches what the tool can actually do, task by task, updated as the tool changes.

Several mechanisms push trust off calibration:

  • Processing fluency. Polished text is easy to read, and what is easy to read tends to be believed.
  • Rare errors are harder to catch. A tool that fails once in 100 cases is harder to oversee than one that fails once in 10, because attention to checking drops.
  • Sycophancy. Chat-based tools tend to agree with whoever is prompting. A tool that agrees with you is not a valid second opinion; real checking uses an independent source.

Calibrated trust is like a good supervisor's trust in a trainee: high for tasks the trainee has shown they do well, low for tasks they have not, and revised as they grow.

Training should show the tool failing, not just succeeding. Usage targets should not be the main goal, because they reward reliance over judgment.

Accountability: Oversight That Is Real

Having a person "in the loop" does not equal oversight. The guidance lists five conditions. The reviewer needs time and attention, expertise to tell a correct answer from a plausible wrong one, information from sources other than the tool, authority to overrule it, and a workplace that does not penalize overruling. Without all five, review becomes a formality.

Two concepts explain the failure. Diffusion of responsibility: when responsibility is shared, each person feels less of it. A moral crumple zone (Elish, 2019) is the pattern where the human closest to an automated system absorbs the blame when it fails, even without the time, information, or authority to prevent the error. The fix: one named owner per important workflow, with real authority and a trained backup, plus a clear rule for disclosing AI use. Red flag: the reviewer approves nearly everything.

Job Design: Redesign the Job, Not Just the Task

The better question is not "Which tasks will AI replace?" but "How will AI change the work?" Searching, drafting, and routine thinking may shrink, while checking, combining information, monitoring, and handling the cases the tool cannot manage may grow. Organizations count the work that disappears and miss the hidden labor that is added.

The job demands-resources (JD-R) model explains why this matters. Job demands (workload, cognitive load) drain energy and lead toward strain and burnout. Job resources (time, support, control) build engagement and buffer those demands. Verifying AI output is a new demand; if schedules do not give it time, it happens after hours. A red flag: reported productivity is up while people describe the work as more draining. Another: AI takes the easy tasks and leaves people with only the hard ones.

Expertise: Protect How Skill Is Built

The same tool affects two people differently. Experts use it to skip effort they no longer need. Novices may use it to skip the very effort that builds skill, so both look equally productive while only one is building the ability to supervise the tool later. This is deskilling through cognitive offloading.

Bainbridge's ironies of automation (1983) describe the trap: automation leaves people with the hardest tasks and a monitoring role, while the routine practice that kept their skills sharp disappears. The out-of-the-loop performance problem (Endsley and Kiris, 1995) is the related finding that passive monitors lose situation awareness and are slower to take over when automation fails.

Skill grows through desirable difficulties (Bjork; see the Memory and Forgetting lesson): effortful practice such as attempting a problem before seeing the answer. So where people are still learning, they should attempt the work first, then compare with the tool's version. Organizations, not individual managers, should decide which tasks people keep doing to build skill.

Fair Process: Introduce AI Fairly

How a tool is introduced affects how people respond to it, regardless of how good it is. This is procedural justice (see the Satisfaction, Commitment, and Stress lesson): fairness of the process, separate from fairness of the outcome. Leventhal's (1980) rules for fair procedures are consistency, bias suppression, accuracy, correctability, representativeness, and ethicality. Employee voice (Thibaut and Walker, 1975) is also central. A fair process does not mean everyone gets what they want. It means input is genuinely considered and can change how, when, or whether a tool is used. Asking for input after the key decisions are made is likely to be detected and may backfire.

Explanations matter. "Efficiency" or "competitors are doing it" explains little; connect the tool to the mission and state its limits. Silence about how roles will change is often heard as bad news, so transparency builds more trust than reassurance.

Problems get reported mainly when people feel safe saying a tool failed or that their own use caused an error. This is psychological safety (Edmondson): a shared belief that the team is safe for interpersonal risk-taking.

Monitoring: Set Limits

AI can collect far more detailed data about how people work. Intensive electronic performance monitoring often falls short of expected performance gains and can increase stress and lower satisfaction and trust. Measuring behavior also changes behavior (reactivity).

Key points from the guidance:

  • Monitoring for safety (such as detecting fatigue) differs from tracking keystrokes to judge performance, and needs different safeguards.
  • Consent is a weak safeguard when the employer is the one asking, because employees may not feel free to refuse. Better tests: Were people told what is collected and why? Is the data used only for a stated purpose? Can people see and challenge what is held about them?
  • Tools that claim to read emotion, engagement, personality, stress, or mental health may be less reliable than they appear. Felt emotions rarely predict behavior, and errors matter when the results affect someone's reputation or job. Such inferences need evidence they are accurate for the intended use and across different groups.
  • Keep data gathered to improve a process out of evaluations of individuals.

When AI Makes the Decision

The guidance does not assume a person must always decide. Research on clinical versus actuarial prediction (Meehl, 1954; see the Clinical Judgment and Interpretation lesson) shows formulas can match or beat unaided human judgment. But accuracy is not the only question: how a decision is made and explained affects whether people see it as fair, and a person responsible for reviewing an AI-supported decision needs real authority to intervene.

Common Misconceptions

Misconception 1: "A human in the loop guarantees oversight."

Reality: Oversight is real only when the reviewer has time, expertise, independent information, authority to overrule, and no penalty for overruling.

Misconception 2: "If a chatbot agrees with my answer, my answer is confirmed."

Reality: Chat tools tend toward sycophancy. Agreement is not a second opinion; check against an independent source.

Misconception 3: "Asking for feedback after the decision counts as a fair process."

Reality: Procedural justice requires voice before decisions are final. Consultation after the fact may backfire.

Key Takeaways

  • APA's 2026 workplace AI guidance is advisory, sets seven human objectives, and excludes personnel selection.
  • Match safeguards to the stakes; one tool used everywhere can repeat one error everywhere.
  • Aim for calibrated trust between misuse and disuse; fluency, rare errors, and sycophancy push toward over-trust.
  • Oversight needs five conditions; without them, blame falls into a moral crumple zone.
  • Checking AI output is real work, and learners should attempt work before seeing the tool's version.
  • Fair process means voice before decisions, honest explanation, and safe error reporting.
  • Limit monitoring to stated purposes, and treat AI inferences about emotions or mental states with caution.

Ready to practice?

Get started in the app