An AI policy template is only worth adopting if each clause targets a known way that artificial intelligence goes wrong at work: people trusting wrong answers, tools inventing facts, confidential information leaking, hidden bias in decisions about people, and missed legal duties such as staff literacy. Below is the evidence for each risk, followed by a template built on it.
Most workplace policies on artificial intelligence rest on three assumptions that rarely get questioned. The first is that keeping a human "in the loop" is a safeguard in itself. The second is that a list of approved and banned uses will shape behavior. The third is that transparency is always good, so asking staff to disclose their use of these tools will build trust.
Each assumption is partly wrong. Decades of research on automation show that people who review machine advice tend to accept it, including when it is wrong, and that experience only partly protects them. Research on rule following shows that a written policy is one influence among many on what employees actually do. Recent experiments show that people who admit to using artificial intelligence are trusted less, which gives staff a reason to hide their use unless the policy deals with that penalty openly.
None of this means a policy is pointless. It means the policy has to be built around how people behave when the tools are in front of them, not around how we would like them to behave. The sections below go through the evidence risk by risk, then turn it into clauses you can adapt.
Why Does the Typical AI Policy Template Fail?
The typical policy reads like a terms of use page. It lists approved tools, forbids entering "sensitive data", tells staff to "use good judgment", and says that a human remains responsible for the output. Every one of those lines sounds sensible. The trouble is that each line assumes a behavior the research says people do not reliably show.
Take "use good judgment". Judgment is exactly the faculty that weakens when a confident machine offers an answer, especially under time pressure. Take "a human remains responsible". Responsibility on paper does not mean scrutiny in practice, and the studies on automation bias, covered below, show the gap between the two. Take "do not enter sensitive data". Staff often do not know what counts as sensitive, and they rarely know what happens to text once it leaves their screen.
There is also a broader question about whether written rules change behavior at all. The best evidence comes from a neighboring field. A 2019 meta analysis pulled together 95 studies on why employees follow or break information security policies. A meta analysis is a study that combines the results of many earlier studies to see what holds up across them. The researchers found 401 different factors that earlier work had tested, which they grouped into 17 categories, and they criticized much of that literature for weak theory. The practical lesson is modest but real: the written rule is one input among many. Whether people comply depends on their attitudes, the norms around them, their confidence that they can do what is asked, and how the rule fits their daily work.
So the gap is this. Most templates describe rules. Few are designed around the specific failure each rule is meant to prevent, and fewer still build in the habits, checks and training that make the rule stick. The rest of this article takes the failures one at a time.
Does Keeping a Human in the Loop Actually Protect You?
The phrase "human in the loop" carries a lot of weight in most policies. The evidence suggests it carries less weight in practice.
Researchers in aviation and medicine have studied this problem since long before chatbots existed. They use the term automation bias for the tendency to accept the advice of an automated system rather than checking it. A 2010 integrative review in a human factors journal described two kinds of error that follow from it. People miss problems the system fails to flag, which the authors call errors of omission, and they act on advice the system gets wrong, which they call errors of commission. The review found these errors in novices and experts alike, in individuals and in teams, and concluded that training and instructions do not prevent them. It also linked automation bias to attention: when people juggle several tasks, they lean harder on the machine.
A 2012 systematic review looked at the same problem in clinical decision support, screening 13,821 papers and keeping 74 that met its standards. A systematic review searches the published research in a structured way so that the conclusions do not rest on a few hand picked studies. It found that heavy workload, complex tasks and time pressure made automation bias worse. It also reported mitigations that seemed to help: training, making users accountable for their decisions, and designing systems that offer information rather than a direct recommendation. Notice the tension with the earlier review, which found training did not prevent the bias. The fairest reading is that training reduces the problem a little but does not remove it, so a policy cannot rely on training alone.
The bias can also outlast the tool. Three 2023 experiments had volunteers complete a simulated medical diagnosis task with help from a system that made a systematic error. Volunteers copied the error while the system was present, and they kept making the same error after the system was removed. In other words, people did not just follow a bad suggestion in the moment. They learned it. These were laboratory studies with members of the public, so the size of the effect in real workplaces is unknown, but the direction is a warning for any organization that lets staff use these tools for repeated judgments.
What helps? A 2021 experiment with 199 participants, published in the proceedings of a peer reviewed computing conference, tested what the authors called cognitive forcing. These are design choices that make people think before they accept advice, for example asking them to commit to their own answer before seeing the machine's suggestion. These designs cut overreliance more than simply showing an explanation. There was a catch. Participants liked the most effective designs least, and the benefit was larger for people who enjoy effortful thinking. A policy that requires an independent view first will work, but it will feel like friction, and some staff will resist it.
The lesson for policy is direct. "A human reviews the output" is not a control unless the policy says how the review happens, when the reviewer forms an independent view, and what the reviewer must check.
How Often Do These Tools Invent Things, and Can Staff Spot It?
Large language models are the systems behind most workplace chatbots. They are trained on huge volumes of text to predict which words are likely to come next, and they produce fluent answers as a result. Fluency is not accuracy. Researchers use the word hallucination for output that sounds plausible but is false or unsupported by the source material.
A widely cited survey, later published in a leading computing journal in 2023, reviewed hallucination across summarizing, dialogue, question answering, translation and other tasks. It concluded that the problem appears across all of them and that no single measure or fix had solved it. That matters for policy because it means hallucination is not a defect of one product that a vendor will patch next quarter. It is a property of how these systems generate text.
Can errors be caught? A 2024 Nature study developed a method that asks a model the same question several times and checks whether the answers agree in meaning. When the answers scatter, the model is probably making something up. The method worked reasonably well across several question sets, but the authors were clear that it cannot catch errors the model makes consistently, such as those learned from flawed training data. The practical takeaway for staff is simple and fits the research: when the same question gets different answers, treat the output as unreliable. When it gets the same answer every time, that is not proof it is right.
A policy therefore needs two things on hallucination. It needs a rule that any factual claim, figure, quotation, citation or legal statement produced by these tools is checked against an original source before it leaves the building. It also needs a rule against leading questions. A question built on a false assumption invites a fluent answer built on the same assumption, and fluency is exactly what makes a hallucination hard to spot.
What Happens to Confidential Information Typed Into a Chatbot?
Most policies say "do not enter confidential information" without explaining why. The research gives a concrete reason.
Language models can memorize parts of the text they were trained on and repeat it word for word. A 2021 security study, presented at a peer reviewed security conference, showed that an attacker who could only send questions to a publicly released model could pull out hundreds of verbatim passages from its training data. These included personal contact details such as phone numbers and email addresses. Some passages appeared only once in the entire training set, and larger models were more vulnerable than smaller ones.
A later study, posted in 2023 as a preprint that had not completed peer review, went further. It extracted gigabytes of training data from open models and from a widely used commercial chatbot. A simple trick made the commercial chatbot release training text about 150 times more often than in normal use. The authors concluded that the safety training these products receive does not remove memorization.
These studies need careful reading. They show that whatever goes into training can sometimes come back out. They do not show that every prompt a user types is used for training. That depends on the product's settings and the contract the organization has signed. The policy point follows from that distinction. If a tool may train on what staff enter, anything entered could in principle resurface for someone else. Employee records, salary data, health information, client files and unreleased financial results do not belong in such a tool. A policy should say which tools have contractual protection against training on inputs, and it should define confidential categories with examples staff will recognize rather than leaving them to guess.
Should Employees Disclose When They Use Artificial Intelligence?
Many policies require disclosure because it seems honest and builds trust. The newest evidence says disclosure is honest but costly, and a policy that ignores the cost will push use underground.
A 2025 paper reporting 13 experiments tested what happens to trust when people disclose that they used artificial intelligence for a task. The tasks ranged from writing and analysis to creative work, and the people disclosing included supervisors, subordinates, professors, analysts and investment funds. Across the experiments, those who disclosed were trusted less than those who did not. The authors traced the drop to perceived legitimacy: observers felt that work done with these tools was less proper. The penalty held whether disclosure was voluntary or required. One finding cuts the other way and matters a great deal for policy. Being exposed by someone else was worse for trust than disclosing yourself.
Four preregistered experiments published in 2025, involving 4,439 participants, found a similar penalty. Preregistered means the researchers recorded their predictions and analysis plan before collecting data, which makes the results more trustworthy. People who used artificial intelligence at work expected to be seen as less competent and less motivated, and observers did in fact rate them lower. These judgments also shaped how observers evaluated job candidates.
Put the two papers together and the dilemma is plain. If a policy demands disclosure but managers quietly mark people down for it, staff will learn to hide their use, which is the outcome that does the most damage when it comes to light. The answer is not to drop disclosure. It is to make disclosure routine and specific, to tell managers what the research shows about this penalty, and to judge work on its accuracy and quality rather than on the tools used to produce it. The penalty in these studies shrank, though it did not vanish, among observers who saw the tools as accurate, which suggests that showing how work was checked may matter as much as saying a tool was used.
Can a Tool That Passes a Bias Test Still Treat People Unfairly?
Decisions about people are where a workplace policy carries the most legal and ethical weight. That includes performance ratings, promotion, discipline, pay, work allocation and dismissal. The evidence here is worrying because the bias can be hidden.
One dialect study, published in Nature in 2024, gave several language models text written either in African American English or in standard American English, with the same content. The models linked speakers of African American English with lower status jobs. In hypothetical criminal trials where the only evidence was a statement from the defendant, the models convicted more often when the statement was written in African American English, about 69% of the time against about 62% for the same statement in standard English. When asked directly about African Americans, the same models gave positive answers. The bias was covert. The authors also found that the training methods used to make models polite and agreeable could widen the gap between what a model says openly and how it behaves.
A 2024 preprint adapted a well known psychology test of implicit attitudes to eight language models that had been trained to avoid bias. The models passed standard bias checks, yet showed stereotypes that mirror those in society across race, gender, religion and health, covering 21 specific stereotypes. The bias showed up most clearly when the models were asked to compare two people directly. This work had not completed peer review when posted, so treat it as supporting evidence rather than settled fact.
Recall the finding from the previous section on automation bias: people who worked with a biased system absorbed its bias and kept it after the system was gone. Combine that with covert bias in the models and the risk to fair treatment is clear. A manager who asks a general chatbot to compare two employees, summarize a disciplinary case or suggest a rating may receive output that is quietly skewed, accept it, and carry the skew into later decisions made without the tool.
The policy response is to separate two kinds of use. Drafting a letter is low risk. Using a general purpose chatbot to judge, rank, rate or compare people should not be allowed at all, and any approved system used in decisions about people should go through a separate review with accountable roles assigned, evidence of testing, and a route for the affected employee to ask for a human review.
What Does the Law Actually Require of Employers?
Legal duties differ by country, but the European Union's Artificial Intelligence Act has become the reference point that many organizations elsewhere use, and two parts of it bear directly on a workplace policy.
The first is the literacy duty in Article 4 of the Act. Under the Act's rules on application dates, this chapter has applied since 2 February 2025. The original text asked organizations that provide or deploy these systems to ensure, to their best extent, a sufficient level of literacy among staff who deal with them. In July 2026 the European Union adopted an amending regulation, often called the Digital Omnibus on artificial intelligence, which softened that wording. As amended, organizations must take measures to support the development of literacy among their staff, without having to guarantee any particular level for any individual. Those measures should reflect the technical knowledge, experience, education and training of the staff, the context in which the systems are used, and the people affected by them. The duty became lighter, but it did not go away, and a policy is the obvious place to record what measures an organization takes.
The second part concerns systems the Act classes as high risk, which include several used in employment. Under the duties for deployers, an organization using such a system must assign human oversight to people with the necessary competence, training and authority. An employer must also inform workers' representatives and affected workers before putting such a system into use at the workplace. Notice how closely this matches the automation bias research. The law does not ask for a human who merely signs off. It asks for a human with the competence and authority to overrule the machine.
Outside the European Union, the most widely used voluntary guide is the risk management framework published in January 2023 by the National Institute of Standards and Technology in the United States. It organizes the work into four functions: govern, map, measure and manage. The govern function is the one most relevant to a policy. It calls for policies that are in place, transparent and actually implemented, for clear roles and lines of accountability, for senior leaders to take responsibility for decisions about these risks, and for an inventory of the systems in use. The framework is voluntary and not a law, but it gives a sensible structure that fits almost any jurisdiction.
An AI Policy Template You Can Adapt
The clauses below turn the evidence above into model wording. Replace "the Organization" with your own entity, adjust the roles to fit your structure, and take legal advice on the law that applies where you operate. Each clause is short on purpose. A policy that staff can read in 10 minutes is more likely to be followed than one they never finish.
Clause 1: Purpose and Scope
This policy sets out how employees, contractors and temporary staff of the Organization may use artificial intelligence tools in their work. It covers tools that generate text, images, audio, code or analysis, whether provided by the Organization or accessed independently, and whether used on Organization equipment or personal devices for Organization work.
The purpose of this policy is to allow staff to gain the benefits of these tools while managing five specific risks: overreliance on output that may be wrong, invented content, exposure of confidential information, unfair treatment of people, and failure to meet our legal duties.
Clause 2: Approved Tools and the Tool Register
The Organization keeps a register of approved artificial intelligence tools. For each tool, the register records what it may be used for, what information may be entered into it, whether the supplier may use our inputs to train its models, and who in the Organization owns the relationship with the supplier.
Staff may use approved tools for the purposes listed in the register. Staff who wish to use a tool that is not on the register must ask the policy owner first. Using an unapproved tool for Organization work is a breach of this policy.
Clause 3: Information You Must Never Enter
Staff must not enter the following into any artificial intelligence tool unless the register expressly permits it for that tool: personal information about employees, applicants or clients, including salary, health, performance and disciplinary records; client confidential material; unreleased financial results; passwords, access codes or security details; and any information marked confidential or covered by a confidentiality agreement.
Staff should understand why this rule exists. Research has shown that these systems can memorize text they were trained on and repeat it to other users. If a supplier may train on what you enter, anything you enter could in principle reappear elsewhere. When in doubt, remove identifying details or do not enter the information.
Clause 4: Checking Before You Rely on Output
You are responsible for the accuracy of any work you produce with the help of an artificial intelligence tool. Before you use or share output, you must check every factual statement, figure, quotation, citation and legal or policy statement against an original source. A citation you have not opened and read must not appear in your work.
For any judgment or recommendation, form your own view before you look at the tool's suggestion, and record that view where the decision matters. Do not ask questions that assume the answer you want, because a question built on a false assumption invites an answer built on the same assumption. If you ask the same question twice and receive different answers, treat the output as unreliable.
Clause 5: Decisions About People
General purpose artificial intelligence tools must not be used to rate, rank, compare, select, discipline, promote, set the pay of, or dismiss any person, or to summarize evidence for such a decision. Only systems approved for that specific purpose by the policy owner and the head of human resources may be used in decisions about people.
Before an approved system is used in decisions about people, the Organization will document how it was tested for unfair outcomes, assign oversight to a manager with the competence, training and authority to overrule it, and inform the affected employees and their representatives where the law requires. Any employee affected by such a decision may ask for it to be reviewed by a person who did not rely on the system.
Clause 6: Disclosure and Attribution
When an artificial intelligence tool has produced a material part of work that you share with clients, regulators, the board or the public, you must say so and state briefly how the work was checked. For internal work, follow the disclosure rules your manager sets for your team.
Managers must judge work on its accuracy, quality and usefulness, not on whether a tool helped produce it. Research shows that people who disclose their use of these tools are often judged less competent, and that being found out is worse than disclosing. The Organization wants honest disclosure, so managers must not penalize it.
Clause 7: Literacy and Training
Before using any approved tool, staff must complete the Organization's basic training on artificial intelligence. The training covers how these tools produce their output, why they can be confidently wrong, the risk of relying on them without checking, how to protect confidential information, and how bias can enter decisions about people. Staff with oversight duties for systems used in decisions about people will receive additional training suited to that role.
The Organization will keep a record of the training each person has completed and will review the content at least once a year, taking account of the roles of staff, the tools in use and the people affected by them.
Clause 8: Accountability, Incidents and Review
The policy owner is a senior manager appointed by the chief executive. The policy owner maintains the tool register, approves new tools, receives incident reports and reports to the executive team on the use of artificial intelligence in the Organization at least twice a year.
Staff must report promptly any case in which confidential information was entered into an unapproved tool, any output that caused or nearly caused harm, and any suspected unfair outcome for a person. Reports made in good faith will not lead to disciplinary action for the person reporting. This policy will be reviewed at least once a year and whenever the law or the tools in use change in a material way.
What Does This Mean for You?
If you are responsible for writing or approving a policy on artificial intelligence, the evidence points to a different starting question. Do not ask which uses to allow. Ask what will go wrong when ordinary, busy, well meaning people use these tools, and write a clause aimed at each failure.
You should expect your reviewers to accept wrong answers, especially when they are under time pressure, so your review rule has to make them think before they see the suggestion. You should expect the tools to produce confident falsehoods, so your checking rule has to state exactly what must be verified. You should expect staff not to know what counts as confidential, so your list has to be concrete. You should expect some managers to think less of colleagues who disclose their use, so your disclosure rule has to address managers as well as users. And you should expect bias in decisions about people to be invisible on the surface, so your policy has to keep general chatbots out of those decisions entirely.
You also need to accept that a document alone will not change much. The research on policy compliance suggests that norms, confidence and fit with daily work matter as much as the text. Your training, your managers' behavior and the tool register will decide whether the policy is followed. Write the policy, then watch how people actually use the tools for the first three months and revise it.
Key Takeaways
- A reviewer who simply signs off on machine output is not a safeguard; experts and novices alike follow wrong suggestions, and the bias can persist after the tool is removed.
- Invented content is built into how language models generate text, so a policy must spell out which claims staff verify and against what.
- Text used to train a model can be pulled back out word for word, which is why a register showing which tools train on user inputs matters more than a vague ban on sensitive data.
- Disclosing the use of these tools lowers trust, yet being caught is worse, so disclosure rules work only when managers stop penalizing honest use.
- Models that pass open bias checks can still produce skewed judgments about people, which justifies keeping general chatbots out of people decisions altogether.
- European law now asks organizations to support staff literacy rather than guarantee it, and it asks for overseers with real authority, not just a signature.
Implications for Practice
Start with the tool register, because almost every other clause depends on it. List each tool your staff use, what it is approved for, and whether the supplier may train on inputs. This single step turns the confidentiality rule from guesswork into something staff can look up, and it meets the inventory expectation in the American risk management framework.
Redesign review around an independent first view. The automation bias research and the cognitive forcing experiment both point the same way: ask people to write down their own judgment before they see the machine's suggestion whenever the decision matters. Expect complaints about the extra step and explain why it exists.
Write a verification checklist for hallucinated content and make it specific to your work. For a human resources team, that might mean checking every statutory reference, every figure in a pay analysis and every citation in a policy paper against the original source. Teach staff to avoid leading questions, since a fluent answer to a question built on a false premise is the hardest kind of error to catch.
Brief your managers on the disclosure penalty. Show them the evidence that people who disclose are rated lower and that exposure by others is worse. Ask them to evaluate work on accuracy and quality, and to treat a clear account of how output was checked as a mark of good practice.
Ban general purpose chatbots from decisions about people and create a separate approval route for any specialized system. Given the evidence of covert bias and of people absorbing machine bias, this is the clause where a firm line is justified. Pair it with a right to human review for any affected employee.
Finally, treat literacy as a measurable program rather than a single course. Record who has been trained, tailor content to roles, and refresh it when tools change. Even under the softened European wording, an organization will need to show what measures it took, and a training record is the simplest evidence.
