On Thursday, Anthropic released reflection, a new feature that lets users track and review their Claude activity. Available now in beta for Free, Pro, and Max users with memory turned on, reflection was designed to help users “reflect on how you use Claude,” per Anthropic.But engineering leaders are skeptical about the feature’s ability to provide truly meaningful insights, and some worry that what’s meant to help users approach AI with more discernment may end up making them even more reliant on AI suggestions and advice. As Sergey Matikaynen, co-founder and CTO of GoGloby, tells The New Stack: “A tool that makes you think about your usage doesn’t improve your judgment.”Reflection tracks usage, not what developers really wantIn the reflection dashboard, Claude users can review their activity over the past 1, 3, 6, or 12 months. A high-level summary outlines what they use Claude for most, including key topics and usage patterns.The feature even highlights specific examples of how people use Claude, “like noting that you often rework email drafts in your own voice, or delegate tasks only after settling the strategy yourself,” Anthropic explains.But what the AI company presents as a catalyst to “reflect on and refine how you use Claude,” some say, fails to deliver enough depth to effect real analysis or improvement. “A tool that makes you think about your usage doesn’t improve your judgment.”For example, when asked whether reflection-style usage dashboards could help developers make better decisions about when to use AI, Anish Agarwal, co-founder and CEO of Traversal and assistant professor at Columbia University, tells The New Stack he sees the potential but doesn’t believe Anthropic’s version captures the right signals: “Knowing I used AI for 80% of my coding doesn’t change how I work.”What he wants, instead, is a system to connect coding behavior to production outcomes to help answer questions like, “Which AI-generated changes led to incidents? Which ones required significant human review? Where did engineers spend hours debugging code they accepted in seconds?” It’s a growing problem for new manager-like engineers who need better signals to evaluate AI’s impact on engineering outcomes. Matikaynen shares Agarwal’s interest in AI-usage insights but agrees that reflection itself is too high-level to be useful. He calls for more outcome-focused insights that capture signals like how long it takes to review AI-assisted pull requests. “If they are delivery signal-based instead of engagement signal-based,” he says, then tracking metrics would be more likely to help developers and other users determine whether to opt for AI or human work. But what Anthropic’s offering up, at least for now, isn’t enough, in his eyes: “It’s a count of activity that has nothing to do with that activity producing anything worth shipping.”Deciding when to use AI requires context AI doesn’t have Anthropic says it designed reflection after interviews with Claude users revealed a desire to better understand how to use AI. Specifically, the AI company says it hopes the new feature will help users answer questions like “How often should someone use AI?” and “When is AI suited to a task, and when is it better left to a human?”To that end, reflection does more than summarize usage; it also provides suggestions on how to use Claude more effectively, “like starting a Project instead of needing to re-explain the context of ongoing work,” says Anthropic. But engineering leaders say answering these questions and determining when and how to delegate work to AI shouldn’t be left up to any vendor or system. When asked where that guidance should come from, Chris Wysopal, co-founder and chief security evangelist at Veracode, tells The New Stack: “AI vendors should explain their models’ strengths and limitations, but they shouldn’t be the final authority on when to use AI.” That decision, he says, should be informed by an organization’s own engineering and security leaders and enforced by an independent governance layer.Moreover, Agarwal points out that AI tools like Claude are, in fact, unable to help users make decisions about AI usage as they lack the greater context and organizational knowledge to make nuanced evaluations: “The model only knows the conversation you’re having with it,” he explains — a drop in the bucket compared to an engineering organization’s years of production history, architectural knowledge, and incident data. “Those are the signals that should determine where AI is appropriate and where human judgment is still essential,” he says.The risk of using AI to judge AI usageAnd then there’s the irony of letting an AI tool prompt users to consider whether they should use AI for a task or rely on human thinking.“Their [Anthropic’s] business model depends on usage increasing, and guidance on restraint from the party who profits from increased usage is always going to be softer than the situation calls for,” points out Sergey.Note that the reflection dashboard itself doesn’t explicitly tell users whether they should or shouldn’t use Claude. Rather, as Anthropic describes it, the feature “invites you to step back and examine the role Claude plays” by occasionally surfacing questions like “What’s one thing you want to keep doing yourself, even if Claude could do it faster?” and offering the chance to talk through the decision with Claude. But Agarwal agrees that Anthropic has a clear incentive to keep users engaged with Claude, which begs the question: Is talking to an AI tool about AI usage even helpful? More worryingly, he cautions about the risk that users will eventually optimize for engagement rather than engineering quality. If the reflection dashboard becomes just another set of numbers to monitor and potentially optimize, it could end up steering users even further away from real engineering goals.The post Anthropic wants you to use AI to decide whether or not you should use AI. appeared first on The New Stack.