A Debate about Managing AI-Assisted Performance Reviews

[00:00:00] Welcome to The Debate. So imagine sitting across from your boss in a small conference room. Oh, we've all been in that room. Right. The classic, slightly too warm office. They hand you your annual performance review. You read it over and wow, it is incredibly insightful. I mean, it pinpoints your exact contributions, it's highly articulate, and it actually gives you a clear, actionable path to a promotion for the next cycle. Which is pretty rare, let's be honest.

[00:00:27] Exactly. So you leave that room feeling fantastic. But now imagine finding out a week later that your boss didn't write a single word of it. An AI did. Do you feel cheated or are you, you know, just glad the feedback was actually useful for once? I think, well, I think the vast majority of people would feel completely betrayed by that. Even if the feedback was technically accurate, you wanted a human verdict. You operated under this assumption that someone actually sat down, thought deeply about your year, and, you know, wrote those words specifically.

[00:00:57] For you. Finding out a machine did all the heavy lifting, it breaks the entire psychological contract of work. But here is the reality we are living in right now. A massive survey from High Wire in August 2026 just dropped a bombshell on the HR world. They found that 78% of managers used some form of generative AI tool to help write feedback in the past year. 78%.

[00:01:18] Yeah, almost everyone. But only 16% of individual contributors were actually told that AI played a role in their review. It's just a massive, completely silent shift in how management works. Which makes it the perfect debate for today, because the asymmetry there is just staggering. It's omnipresent for managers trying to save time and completely invisible to the employees whose actual careers are on the line.

[00:01:39] Right. So the core question we are exploring today is whether AI-assisted performance feedback fundamentally enhances managerial capability and the quality of feedback or... Or if it inevitably erodes the trust, the fairness, and that authentic human judgment that people need to actually grow in their careers. Exactly. And I am taking the position that when AI is properly governed to preserve human ownership, like when we use it for preparation and drafting, it's a massive upgrade.

[00:02:06] It takes a dreaded, often poorly executed administrative chore and turns it into precise, genuinely useful developmental feedback. And I am arguing that outsourcing the synthesis of a human being's performance to an algorithm inherently destroys interpersonal trust. It actively makes hidden systemic biases worse, and it degrades the very coaching skills managers are supposed to have in the first place. The whole point of feedback rests on the recipient believing a human observed them and cared enough to weigh their work.

[00:02:34] I mean, generative AI just shatters that belief. Well, let's start by looking at the practical reality of being a manager today. Historically, managers are fairly terrible at performance reviews. Universally. It's a known fact. Right. It's viewed as this heavy, compressed administrative burden at the end of the year. People are overwhelmed. But it's not just a workload issue, is it? There's a profound skill gap. Most managers struggle immensely to deliver constructive criticism. Yeah, they do. They get anxious about employee reactions, so they sugarcoat things.

[00:03:04] Or they write incredibly vague, unhelpful notes. I agree that managers struggle, but handing them a ghostwriter doesn't fix the underlying lack of management skill. It just, you know, just papers over it. It actually does fix it, though, if you look at the data. Let me break down a field experiment by Tong and colleagues from 2021. They isolated something they called the deployment effect. They found that when managers use AI to generate feedback, employee performance actually improves afterward. Wait, why would it improve just because a machine wrote it?

[00:03:31] Because the output is far more accurate and data-rich than what a stressed-out supervisor writes at 11 o'clock on a Friday night. Human managers get emotionally fatigued. The AI doesn't. Sure, but... Let me just finish this thought. The AI bridges the gap between what a manager intellectually knows they should say and what they can actually articulate on paper. If organizations embrace AI as a rehearsal and drafting partner, it's a tremendous net positive for the employee who finally gets clear direction.

[00:03:59] See, you are focusing entirely on the mechanical output, the fluency of the words on the page. And in doing so, you are completely ignoring the relational mechanism of feedback. What do you mean by the relational mechanism? I mean why someone actually listens to feedback in the first place. There was a massive research project by Shilk and Ryman in 2025. They ran 13 separate experiments to map out something we all intuitively feel. They found a severe trust penalty when AI is used in relational tasks. A trust penalty just for using the tool?

[00:04:29] Even if the tool produces a better result? Exactly. Because audiences expect human effort in work that signals care and judgment. Look, if I am getting my car repaired, I don't care if an AI diagnoses the engine. I just want it fixed. But if my career is being evaluated, I expect human effort. When you use AI to write a review, you aren't just getting better grammar. You are triggering automation bias. Okay. Well, let's unpack automation bias because I hear that term thrown around a lot in these discussions.

[00:04:57] It's a foundational concept from human factor psychology. Per Surman and Manzi proved back in 2010 that human beings inevitably over-rely on automated recommendations. We get lazy. If the computer gives us a polished, friction-free paragraph that sounds professional, we just rubber stamp it. We stop thinking critically. But that assumes... Mangers will cede their most important leadership duty, evaluating their people, to a language model because it is the path of least resistance. I see the risks there. I really do. But let me push back on this trust penalty idea.

[00:05:26] The problem isn't the AI itself. The problem is the secrecy. The reason there is a trust penalty is because employees feel tricked. Well, yeah, because they are being tricked. Right now, yes, when it's hidden. But what if we change the rules of engagement? I propose disclosure by default as the solution to this trust gap. Disclosure by default? Yes. Right now, because we don't have shared corporate norms, an individual manager admitting they used AI feels like they're confessing to taking a lazy shortcut.

[00:05:55] But look at what JP Morgan did recently with their internal guidance, or Meta rolling out their AI performance assistant. They are creating standard shared expectations across the whole company. And you think telling employees, hey, a machine wrote your review, suddenly makes it okay? If the organization openly sanctions it, yes. It stops feeling like a manager's personal failing and starts signaling that the manager is just complying with a modern, legitimate corporate standard. Transparency neutralizes the betrayal. Wow.

[00:06:24] That is a highly optimistic reading of human psychology. And frankly, the data does not support it at all. Even if you disclose it, the psychological realities of AI-mediated communication are incredibly unforgiving. How so? I mean, if everyone knows it's happening and it's out in the open, where's the harm? Hohenstein ran experiments in 2023 showing that just the mere suspicion of AI being involved changes how messages are received.

[00:06:49] When people know AI is in the loop, the sender, in this case the manager, is rated as less cooperative. It kills the feeling of affiliation. And let's look back at that very same Tong study you cited earlier to praise the deployment effect. Right, the one where AI feedback improved performance. Yes, but you left out the second half of their findings. They discovered a massive catch-22. They called it the disclosure effect. The moment employees actually learned the feedback came from AI, it actively reduces their performance.

[00:07:19] Wait, the exact same feedback? The exact same feedback. The technical quality did not change at all. But knowing an AI wrote it caused employees to view the feedback so negatively that their performance dropped. And the underlying psychology here is fascinating. It goes back to Kluger and Denise's feedback intervention theory. Which says what exactly in plain English? It says that feedback only works when it directs a person's attention to the task itself and how to learn from it.

[00:07:46] But when a review triggers suspicion about who actually wrote it, it pulls the employee's attention away from the work and entirely onto themselves and their relationship with their boss. Interesting. Yeah. They stop asking, how can I improve this project? And start asking, does my manager even care about me at all? Okay. That makes sense on an emotional level. I'll give you that. But let me offer an analogy to challenge this idea of authenticity. Corporate executives have used human speechwriters for decades.

[00:08:13] HR departments have competency libraries with hundreds of boilerplate phrases that managers just copy and paste from every single review cycle. Well, that is true. Right. And we do not view a CEO as inauthentic just because a speechwriter drafted the exact phrasing of their keynote, as long as the CEO takes accountability for the final words. So, if a manager uses AI to organize their rough notes, synthesize the data, and generate a draft, but they ultimately sign their name and own the evaluative judgment,

[00:08:42] why is an AI ghostwriter uniquely destructive compared to a human ghostwriter or an HR template? Because an AI is not just a speechwriter cleaning up your dictated thoughts. A speechwriter takes a CEO's core ideas and polishes them. Generative AI is an engine of synthesis. Managers aren't just doing light copy editing here. They are letting the AI look at a year's worth of an employee's Slack messages, Jira tickets, and emails and asking it to synthesize a professional verdict. But the manager is still in the loop.

[00:09:11] They still have to read it and approve it before it goes out. That is the illusion of the human in the loop. We talked about automation bias earlier, and this is where it gets incredibly dangerous. When an AI crafts a highly persuasive, well-written narrative, managers will sign their names to conclusions they never actually form themselves. Look at IBM's experience. What happened at IBM? They had a machine learning compensation system a while back. The system looked at the data and recommended salary increases. But the final decision was completely left up to the human managers.

[00:09:40] They were explicitly told, you make the call. Do you know how many managers actually disagreed with the AI's recommendations? I'm guessing not many if you're bringing it up. Fewer than 5%. 5%? 5%. That is not a human exercising independent judgment. That is a human acting as a legal liability shield for a machine. And when you apply that to performance reviews, you get two massive problems.

[00:10:03] You get omission errors, where an employee's quiet behind-the-scenes contributions are entirely missed because the AI didn't happen to scrape that data. Right. And you get commission errors, where the manager passively nods along to a false critique just because the model suggested it in a very authoritative tone. I hear you. And yes, a 5% disagreement rate sounds terrifying. But I am not convinced that this is an inevitable outcome of the technology itself. What you are describing is a software design problem.

[00:10:30] It's a UI issue, rather than a fundamental flaw in AI. How is surrendering human judgment just a UI issue? Because it depends entirely on how the software prompts the human. If the interface simply spits out a perfectly formatted three-page review and just asks the manager to click a big green approve button, then yes, you get automation bias. People are busy. They will click approve. But modern platforms are specifically engineering against that exact behavior. Unskeptical. How do you engineer out human laziness?

[00:11:00] By forcing judgment before drafting. Take Lattice's evidence-based review drafts, which launched in July 2026. This is the ideal structural framework. The AI agent can help draft answers, but it is strictly configured so it cannot propose or set the actual performance rating. So the manager still has to pick the number. Exactly. The human manager must independently form and input the rating and bullet out their core observations first. Only then does the AI help draft the narrative text based strictly on those bullets.

[00:11:28] And if the manager hasn't provided enough evidence, the AI actually flags them and says, hey, there is not enough data here to support this rating. If we structurally force human judgment before the algorithmic drafting, we neutralize the automation bias you were so worried about. Okay. Restricting the AI from selecting a number from one to five is a good safeguard. I will give you that. But we still have to grapple with the narrative text the AI generates based on those bullets. The text is just reflecting the manager's bullets.

[00:11:55] But it is expanding on them using a model trained on historical human data, which means you are institutionalizing historical bias at an unprecedented scale. There is robust, deeply documented research from Carell and Samard showing that women already receive vaguer or less actionable feedback than men in professional settings. Yes, that's true. Men get actionable business feedback. Women often get vague praise or criticisms about their communication style or personality. Right. That is a known longstanding human flaw in management.

[00:12:22] And generative models exacerbated exponentially. There was a major peer-reviewed study by Juan and colleagues in 2023 and another analysis by Snyder that same year. They looked at what happens when AI generates evaluations. The AI actively infers gender from job titles or pronouns and then it subtly adjusts its output. It often writes longer and significantly more critical feedback for women. So the AI is quietly reproducing the same systemic issues human HR departments have been fighting for decades. Exactly.

[00:12:49] A manager types in, write a balanced review, thinking the machine is perfectly objective. But the machine is inherently biased by its training data. It bakes those hidden inequities right back into the official record. I understand the gravity of that, truly. But I want to flip this paradigm entirely for a second. Instead of viewing AI only as a biased text generator, why not look at it as our most rigorous, scalable auditor? An auditor? Yes. Human biases have plagued performance reviews for a century.

[00:13:17] As you just pointed out, humans are terrible at this. AI is actually the only tool we have that can audit these biases at a massive, company-wide scale. But how does an inherently biased model audit for bias? By being specifically prompted and designed to do so. Consider SAP's late 2024 updates. They added a text analyzer specifically designed to scan human written reviews before the managers allow to submit them. So it reads over their shoulder? Exactly. It interrogates the text.

[00:13:45] It actively looks for vague praise, unsupported personality labels, and gender descriptors. If a manager writes that a female employee is abrasive, the AI flags it and suggests focusing on specific communication outcomes instead. So rather than a manager sitting alone in their office with their unconscious biases, they have an active calibration partner. If we use AI to audit the language and mandate that feedback is tied to evidence, we are actively improving organizational justice.

[00:14:12] Okay, I will concede that using AI to audit human text is far safer than using it to generate text from scratch. But we have to look at the broader institutional pressures here. Organizations are not just using AI defensively to catch bias. They are aggressively pushing its adoption to cut corners. And its influence on workers' livelihoods has crossed a dangerous threshold that we just can't ignore anymore. And that is exactly why the incoming wave of regulation is going to force healthy governance. We aren't operating in the Wild West anymore.

[00:14:41] If you are managing a team right now, the legal landscape is shifting under your feet. You mean the EU regulations? Yes, the European Union's AI Act specifically classifies AI used to evaluate workers as high risk under Annex 3, with a hard compliance deadline in December 2027. But it's not just Europe. In the U.S., Colorado's SB 26-189 takes effect in early 2027. These legal frameworks require human oversight. You have to keep mandatory system logs. You have to notify employees.

[00:15:10] You have to provide formal appeal routes. So it just becomes a massive compliance checklist. It becomes a legal paper trail that prevents the lazy behavior you are worried about. By adhering to these strict frameworks, companies can safely use AI to completely eliminate the sheer administrative burden of review writing. And here is the ultimate payoff for the employee. Companies can redirect all those saved managerial hours into continuous face-to-face coaching. We make the annual review lighter so the day-to-day human development can be heavier.

[00:15:38] I'm sorry, but the idea that saved administrative time will magically be reinvested in human connection is a total corporate myth. Why is it a myth? If a manager gets 10 hours back a month, why wouldn't they use it to coach their team? Because the data completely contradicts it. Look at WTW's 2026 Employee Experience Global Market Study. They surveyed 7,000 employees. The percentage of employees who say their manager actually helps develop their abilities drop from 84% to 78% in just four years.

[00:16:07] AI is making it easier for managers to skip the hard, uncomfortable work of genuine leadership. But surely clear regulations will curb the worst of that. Regulatory compliance won't prevent the cultural rot of treating employees like data points. Companies like Shopify and Microsoft are making AI use of core performance expectation. Metagrades workers on AI-driven impact. When you incentivize managers to use AI everywhere, they will use it to write reviews, regardless of the compliance checklists HR puts in place.

[00:16:35] And when governance fails, the consequences are incredibly real. Look at the Deliveroo case in the Italian labor court. The delivery company? Yes. They got hit with a 50,000 euro fine because their blind algorithmic evaluation system couldn't account for basic human context. It was penalizing workers for not logging on, without understanding they were sick or striking. When 15% of workers already believe AI gets more room to correct mistakes than they do, outsourcing their feedback to a machine just breaks the fundamental contract of the workplace. Well, we have covered incredible ground today.

[00:17:06] To summarize my position, AI is already deeply embedded in the modern workplace. We simply cannot put the genie back in the bottle. But by leaning into structured, evidence-based tool design systems like Lattice that strictly separate gathering evidence from the final evaluative judgment, we can mitigate the risks. If we use AI as a rigorous bias auditor and a private rehearsal partner for managers, organizations can elevate performance reviews.

[00:17:31] We can transform them from a dreaded administrative chore into a precise, data-rich tool that actually serves the employee. And to summarize my stance, the inherent cost to interpersonal trust is simply too high. The subtle but incredibly powerful pull of automation bias means managers will inevitably surrender their critical judgment to the machine. Combined with the risk of reproducing hidden systemic biases, outsourcing the synthesis of a human being's hard work to a language model degrades the relationship between manager and employee.

[00:18:01] Feedback only matters if the recipient truly believes they were seen by their leader. I think where we fundamentally converge today is on the danger of doing this in the shadows. We both agree that the unreviewed, hidden use of public generative AI tools by managers, where confidential employee data is just dumped into a public prompt and an unedited review is copy-pasted back, is an incredibly dangerous experiment. Absolutely.

[00:18:25] The high-wire data showing so many employees suspecting undisclosed use represents a massive ticking time bomb for corporate culture. It threatens organizational justice, data privacy, and just basic trust. The source material we've explored today offers a complex, shifting landscape. It leaves us to consider whether the efficiency, the polish, and the potential objectivity gained from AI are ultimately worth the risk of eroding authentic human judgment in our own workplaces.

[00:18:53] It brings us right back to our original question. Are we building tools that give a manager a better magnifying glass to see their employees work more clearly? Or are we building a machine that simply prints out a certificate, completely bypassing the human eye? That is the question every leader must answer. Thank you for joining us on The Debate. Thank you.