Ghostwritten Feedback: Managers' Hidden Use of AI in Performance Conversations, and How Leaders Can Protect Trust and Development

[00:00:00] So picture an analyst sitting at her desk, right? She's staring at her year-end performance review. Okay, we've all been there. Exactly. And the paragraphs on the screen are, you know, they're polished, beautifully balanced, and just oddly frictionless. Right. The praise is warm, but it feels kind of generic, like it could describe almost anyone in her department. Yeah, that classic corporate speak. Right. Then she gets to the development section. And it recommends she, quote,

[00:00:26] continue to build stakeholder influence, which sounds like great actionable advice, right? Sure. Yeah. But the problem is, she saw that exact same phrase in a colleague's review earlier that week. Oh, wow. Yeah. So she sits there looking at this document, and she can't point to a single factual lie. I mean, the metrics are correct. But something's missing. Exactly. She can't find her manager's voice anywhere in the text. And that tiny lingering seed of doubt is planted, like, did my manager even write this? Yes.

[00:00:56] And that moment of doubt, that's, it's rapidly becoming a structural feature of the modern working world. It really is. I mean, it's moved far beyond a few isolated anecdotes and into the absolute core of how businesses operate today. Which brings us to the mission of our deep dive today. We are unpacking this really fascinating 2026 paper from the Human Capital Leadership Review. By Dr. Jonathan H. Westover, right? Yes.

[00:01:20] Titled, Ghost Written Feedback. And basically, we're going to explore what actually happens when the most deeply personal part of management, which is, you know, evaluating another human being's performance, is just quietly outsourced to generative AI. Because we really need to figure out how to use this tech without completely destroying the foundation of workplace trust. Exactly. And just to ground us here, how big of an issue is this really?

[00:01:45] Well, to understand the sheer scale of this phenomenon, we have to look at the August 2026 High Wire Survey. They surveyed over a thousand U.S. corporate professionals and the numbers are just staggering. What did they find? So, among the people managers, a massive 78% admitted they had used an AI tool to help write employee feedback or a performance review in the previous 12 months. Wait, 78%? 78%.

[00:02:11] That is essentially four out of every five managers relying on software to evaluate their teams. It is. But the more critical data point is on the receiving end. Among the individual contributors who received those reviews, only 16% were told that AI played any role whatsoever in the drafting process. Wow. Only 16%. Yeah, there is this massive silent gap. The technology is practically everywhere in performance conversations.

[00:02:38] But it remains entirely invisible to the people whose careers are actually being evaluated. So think about your own recent performance reviews. When you were reading through that document, absorbing the criticism, feeling good about the praise, was it truly written by a human being who spent time thinking about your year? Or was it ghostwritten? Right. So we know this is happening at a massive scale, but we really have to look at the forces driving it.

[00:03:04] Because we are moving away from individual managers just looking for a shortcut, and we are entering an era of massive corporate mandates. Yeah, there are basically three distinct forces pushing this adoption. And it goes far beyond the cliche of the lazy boss. Okay. So what's the first one? The first driver is simply workload. I mean, performance management has historically been this heavy, highly administrative ritual. Right. It's a huge burden.

[00:03:30] Exactly. Instead of continuous coaching throughout the year, everything gets compressed into a few highly stressful weeks in November or December. So if a piece of software can turn a scattered list of bullet point notes into a cohesive, polished paragraph in three seconds? The appeal for an overworked manager is totally undeniable. It functions as the ultimate stress relief valve for a system that was already bottlenecked. Makes total sense.

[00:03:56] But time is only one part of the equation here, right? Because the actual act of writing the review is an emotional hurdle. A massive hurdle. And that brings us to the second driver, which is the skill gap. Delivering constructive criticism is just a genuinely difficult human skill. Yeah, nobody likes doing it. No, they don't. In fact, the source material highlights a Gallup survey showing that only 14% of employees feel inspired by their reviews to improve. Only 14%? Right. That's abysmal.

[00:04:26] Right. Managers know their track record is poor. They struggle to anticipate how an employee will react. They fight to stay objective. And they often just lack the vocabulary to put very nuanced, subtle observations into words. So generative AI steps in with this promise to bridge the gap. Exactly. The gap between what a manager wants to communicate and what they can actually comfortably write down. Okay. So the manager feels inadequate. The AI feels confident.

[00:04:53] But you mentioned a third driver, and this one changes the entire context. Yeah, the push isn't just coming from stressed out middle managers anymore. It's coming straight from the C-suite. Wait, really? The executives are mandating this. We are seeing heavy institutional signaling from the very top. Look at the corporate directives that came out in 2025 and 2026. Shopify's leadership actually told employees that reflexive AI use is a baseline expectation. Oh, wow. Yeah.

[00:05:22] And at Microsoft, leaders in the developer division explicitly told managers that utilizing AI is now core to every role and should actively factor into how people are assessed. So it's not even optional anymore? Nope. Meta even rolled out an internal AI performance assistant, specifically designed to help employees draft performance content. That is wild. So if you are a manager in one of those environments, your own end of year rating might actually depend on how heavily you lean on these tools.

[00:05:51] Using AI to write the reviews becomes a performance metric in itself. Exactly. But there is a glaring disconnect here. I mean, it reminds me of having an incredibly eloquent, emotionally intelligent friend write an apology letter to your spouse on your behalf. Oh, that's a great way to put it. Right? Yeah. Like the grammar might be absolutely perfect. The structure might be flawless. It might even identify the root cause of the argument better than you could. But it completely misses the point of an apology. Exactly.

[00:06:20] The entire value of that communication relies on the recipient believing that you paid attention, that you did the emotional labor, and that you chose those words specifically for them. That analogy perfectly captures the core issue at play, which researchers call the trust paradox of AI feedback. The trust paradox. Yeah. Here is the tension. AI actually does help managers write clearer, more actionable, and more data rich reviews. The output is often objectively better. So why is it causing a crisis of trust?

[00:06:49] Oh, wait, hold on. Let me push back on that for a second. Sure. If an AI can comb through a whole year of my Slack messages, Jira tickets, and emails, and produce a review that is factually more accurate and data rich than what a stressed out manager could possibly piece together at 11pm on a Friday, why should I care? You're saying just focus on the results. Yeah. If the data points are correct and they help me secure a promotion or a raise, shouldn't I just put my ego aside? Why does the emotional origin of the feedback even matter?

[00:07:17] See, that sounds completely logical on paper, but it fundamentally ignores human psychology. Behavioral scientists have run extensive field experiments like this fascinating 2021 study by Tong and colleagues separating what they call the deployment effect from the disclosure effect. Okay. Break that down for me. So when managers deploy AI, the feedback generated actually does improve employee performance, largely because of that exact data richness you just mentioned.

[00:07:46] So it works? It works. Right up until they tell the employee. The moment managers disclose to the employee that an AI wrote the feedback, the employee's performance actively drops. Wait, what? Yeah. It drops? Yes. They start viewing the feedback far more negatively, even though the actual text hasn't changed at all. So the feedback is highly effective right up until the second I find out a machine wrote it, at which point it actively harms my performance. What is happening in the brain there?

[00:08:16] We can understand the mechanics of this by looking at how humans process criticism. There's decades of psychological research on feedback intervention, specifically a major meta-analysis by Kluger and Denise. And what do they find? They show that effective coaching only works when it directs the recipient's attention to the task itself and to learning. You read the review and you think, okay, how do I improve my coding speed? Right. You focus on the work. Exactly.

[00:08:44] But when an employee suspects their manager used a machine to generate the review without telling them, their attention violently shifts away from the task. It shifts to the self and to the relationship. Oh, I see. They stop asking, what should I do differently? And they start asking, does my boss even care about my career? Am I just a data point to them? Because the legitimacy of the evaluation isn't just about the facts. It's about the social contract. The social contract is everything.

[00:09:11] Studies by Shilke and Reiman, they ran 13 different experiments on this exact dynamic. They show that legitimacy and trust are inherently social constructs. So we need that human element. We do. Audiences demand a signal of human effort in any work that requires judgment or care. When that effort is outsourced, the feedback just loses its social legitimacy. That makes a lot of sense. And it goes even deeper. Just the mere suspicion that AI was used changes how a message is received.

[00:09:40] Even if a manager wrote every single word by hand, if the employee suspects they used the chatbot, the employee begins to view the manager as less cooperative and less trustworthy. Oh, wow. So it creates this paranoid work environment. And the source text points out this isn't just damaging for the employee on the receiving end, right? The managers themselves are suffering. Yeah, they suffer cognitive decay through a phenomenon called automation bias. This was studied by Parasaraman and Manzi. Automation bias. What does that look like in practice?

[00:10:09] It happens when people start over-relying on automated decision aids. It leads to two very specific types of managerial failures. Errors of omission and errors of commission. Okay, so what's an error of omission? An omission error occurs when the AI fails to flag a recurring problem or entirely misses a massive contribution an employee made, and the manager just blindly accepts the summary. They just assume the AI caught everything. Right.

[00:10:35] And a commission error is when the AI hallucinates or suggests a weakness that isn't really accurate, and the manager accepts it without verifying the claim. Oh, that's dangerous. So they end up signing off on conclusions they never actually formed in their own minds? Exactly. Because the physical act of writing feedback forces a human being to retrieve memories, weigh conflicting observations, and, you know, commit to a viewpoint. It forces you to actually manage. Right. If you outsource the synthesis to a language model,

[00:11:04] you lose the observational habits that made you a competent manager in the first place. You just stop paying attention because you know the machine will summarize it later. Which ultimately flattens out the real differences between employees on a team. Think about it. If every manager in a department uses the same language model, relying on the same underlying algorithms, all the performance reviews start converging on the exact same vocabulary and tone. So everyone sounds identical. Pretty much. To an executive skimming the files, it looks like wonderful organizational consistency.

[00:11:34] But in reality, calibration becomes impossible. The distinct human verified signals of truly exceptional or truly struggling employees have just been smoothed over by the software. So we know the human touch matters for building trust, and we know managers risk losing their evaluative edge. But let's look at the actual content these models are generating. What happens to fairness when we hand the evaluation pen over to a machine? That is a huge question. Because you'd hope that using a neutral computer program would strip out all our human biases.

[00:12:04] I mean, a machine doesn't care about gender or race, right? That is the absolute most dangerous assumption a company can make. Let's look specifically at gender bias in performance reviews. Okay. Human managers already struggle deeply with this. Decades of research, like the work by Carell and Samard, show that women are systematically less likely to receive specific, actionable feedback tied to business outcomes. So what do they get instead?

[00:12:30] They are far more likely to receive vague praise or criticism centered entirely on their communication style or personality compared to their male counterparts. Wow. So the historical archive of human written reviews is already deeply flawed. What happens when the AI takes over? It exacerbates those flaws significantly. Because these large language models are trained on that exact historical data, they have learned our biases perfectly. Oh, no. Yeah.

[00:12:57] Researchers like Juan and colleagues and Snyder have found that language models often infer an employee's gender simply from their job title or department. Just from the title. Yep. And when prompted to write feedback, the models will systematically write longer reviews for women. But that extra length isn't filled with helpful metrics. It's typically filled with more criticism, vague personality observations and gendered assumptions.

[00:13:23] So asking an AI to quote, write a balanced review is not like pouring dirty water through a filter to clean out the impurities of human bias. Not at all. It's more like walking into a funhouse mirror. It takes the exact same societal biases we already possess, reflects them, distorts them and makes them larger. That's a perfect analogy. But the terrifying part is that it wraps those biases in a highly authoritative, confident, frictionless tone that makes the manager think the review is perfectly objective.

[00:13:51] And this moves the conversation from just an HR problem to a massive legal liability. We are standing on the edge of a regulatory cliff regarding automated employment decisions. We really need to talk about the laws coming down the pipeline because this is going to catch thousands of companies completely off guard. The landscape is shifting rapidly. Over in the European Union, the Comprehensive AI Act explicitly classifies AI systems used to evaluate workers as high risk. High risk.

[00:14:20] What does that actually mean for a company? That classification triggers a cascade of strict obligations. We're talking mandatory transparency, rigorous data logging and guaranteed human oversight. And enforcement for that kicks in by December 2027. Okay, so the EU is moving fast. But what about the US? Are state laws already jumping on this too? They absolutely are. Colorado recently passed Senate Bill 26-189, which becomes effective in January 2027.

[00:14:50] It requires employers to give workers explicit notice if AI materially influences consequential employment decisions. And does it give employees a way to fight back? Yes. It mandates a clear route for human review if an employee faces an adverse outcome. And we already have stark warnings from the international courts about what happens when algorithms evaluate humans without context, right? We do. A landmark example is the 2020 Deliveroo case in Bologna, Italy.

[00:15:19] Deliveroo, the food delivery company, used an algorithm to rank its riders based on a metric called reliability. Okay. Riders with higher reliability scores were granted better access to the most lucrative shifts. But the algorithm was blind to human context. Oh, so? It could not distinguish between a rider who missed a shift due to a documented illness or a legally protected labor strike versus a rider who simply blew off the shift because they didn't feel like working. Oh, wow.

[00:15:47] It just saw a missed shift and penalized them equally. Exactly. The court ruled the system inherently discriminatory. It's a vital lesson for any performance system. An evaluation tool that cannot account for protected human context will treat legally valid absences and poor performance as exactly the same thing. The reality is the technology is out there and organizations cannot simply ban it.

[00:16:12] We already know that managers will just secretly paste confidential employee data into public AI tools, which creates an absolute nightmare for data privacy. A total nightmare. Yes. So if you are leading a team right now, how do you use these tools to enhance human connection rather than replace it? The source material lays out a very clear evidence-based playbook for this. And the first and most crucial step is to change the culture around disclosure. Organizations need to disclose by default. Well, wait.

[00:16:41] Didn't we just establish earlier that disclosing AI used to an employee drops their trust and lowers their performance? We did. However, behavioral research shows that being exposed by a third party or an employee figuring it out on their own is far more damaging to the relationship than upfront self-disclosure. Ah, so it's about getting ahead of it. Right. Given that nearly half of employees already suspect their managers are secretly using AI, the choice isn't between disclosure and blissful ignorance.

[00:17:09] It is between disclosure on the manager's terms or discovery on someone else's. That makes sense. The strategy is to make AI use a shared, transparent organizational norm rather than an individual confession of laziness. So if a company like Meta or J.P. Morgan officially sanctions a specific internal AI tool for the review cycle, it changes the meaning of the disclosure. Exactly.

[00:17:35] When the company makes a policy, a manager admitting they use the tool stops signaling a lack of care. It starts signaling compliance with a known secure corporate standard. Which leads directly to the next safeguard. Keeping the actual judgment human. Even with an official policy, you have to prevent managers from just hitting auto-generate on a final score. How do you do that? Well, platforms like Lattice have introduced features like evidence-based review drafts.

[00:18:05] The AI is allowed to pull from existing data and draft performance summaries, but it is hard coded to prevent it from ever selecting the actual performance rating. Oh, that's smart. Yeah. The human manager is forced to evaluate the evidence and make the final judgment call themselves. So you are basically building friction back into the system on purpose to force the human to think. Exactly. But if they aren't using the AI to write the final verdict, where does the AI provide the most value?

[00:18:34] Its highest value is in rehearsal, not substitution. Managers should use the AI as a sparring partner before the actual performance meeting. Like role-playing. Yeah. You can prompt a language model to play the role of an upset defensive employee, allowing you to practice your responses to difficult pushback. Interestingly, the high-wire data shows a massive asymmetry in who gets access to this. What do you mean?

[00:19:00] Half of all managers use AI to practice for high-stakes situations, but only about a third of individual contributors do. Oh, companies should absolutely be giving these rehearsal tools to the employees as well. Imagine an employee being able to practice asking for a raise or role-playing how to receive critical feedback without getting defensive before they ever step into the boss's office. It would level the playing field so much. But what about the funhouse mirror of bias we talked about?

[00:19:27] How do we use the tool without baking in those historical prejudices? You flip the role of the AI. Instead of using the language model to write the initial draft, you use it to audit human-written text for bias. Okay, how does that work? Enterprise software companies like SAP have added text analyzers that monitor a manager's writing in real time. If a manager writes a review filled with vague praise or heavily gendered descriptors, the AI flags it before it's ever submitted. So it acts as a coach for the manager.

[00:19:58] Right. Forcing the manager to revise their language to be more specific, metric-driven, and inclusive. Okay, let me pause and ask you a crucial question about all of this. Sure, go ahead. Let's say a company does everything right. They buy the secure enterprise tools, they force the managers to make the final ratings, they use AI to audit for bias, and they use it to rehearse. As a result, the AI saves a manager three hours of tedious typing during review season. Okay.

[00:20:25] Isn't there a massive risk that the manager just takes those saved three hours and uses them to fill out other administrative paperwork instead of actually spending time coaching their team? You have just hit on the absolute central warning of Dr. Wistover's research. All of this technology is merely a lever. If AI saves a manager three hours, leadership has to ruthlessly track where those hours are reinvested. Because they don't?

[00:20:49] Well, global workplace studies show that the number of employees who feel their manager actively helps develop their abilities has been steadily dropping for years. Coaching is eroding at the exact historical moment that AI is making it incredibly easy to fake it. The time saved on administration has to be converted into human interaction. It has to be. The saved time must be reinvested into frequent, shorter human conversations throughout the year.

[00:21:13] If a manager gets all their performance reviews submitted in record time, but their employees rate the feedback is generic and feel they haven't been coached, that manager has a fundamental developmental failure. Wow. No amount of perfectly generated text will fix a broken human relationship. So really the core lesson of this deep dive is that AI is a phenomenal instrument for gathering scattered evidence, for auditing our flawed human biases, and for rehearsing difficult conversations.

[00:21:43] Absolutely. But the actual judgment, the verdict on someone's career trajectory and the interpersonal delivery of that verdict must always carry a human signature. The boundary between machine fluency and human judgment has to be fiercely protected. So if you are managing a team right now and you sit down to write a review tomorrow, or if you are the one sitting in the chair receiving that review, you now know exactly where that invisible boundary lies. Yeah.

[00:22:11] You know the profound difference between a tool that helps you see your employee more clearly and a tool that you use to avoid looking at them at all. It really forces us to ask what kind of long-term culture we are building with these systems. Which brings up one final, incredibly unsettling thought I want to leave you with. It builds on a data point from that High Wire survey. Oh, the one about mistakes? Yeah. They found that 15% of workers currently believe AI gets more room to correct its mistakes than human employees do.

[00:22:41] That's a chilling thought. Think about the implications of that. If organizations start treating the errors an AI makes as just harmless iterations needed to improve the model, but they treat an employee's errors as permanent, unforgivable data points fed into an evaluative algorithm, what kind of workplace are we actually building? Right. Are we inadvertently creating a culture where human beings learn to hide their mistakes entirely?

[00:23:08] Because if a flawless machine is always watching, always summarizing, and always writing your file with perfectly frictionless fluency, the absolute safest thing for you to do is to never take a risk at all. I'm going to read the lesson about the next one, and I will, as I take my first. Well, as I have to remember, we won't be able to apply this up to it. I'll be able to ask them about the fourth line, but I'm going to fail. Thank you.