{
  "source_file": "0930",
  "source_platform": "Ghost",
  "author": "Louis-François Bouchard / What's AI",
  "title_listed": "How to Control Bias in AI Agents",
  "keywords": "",
  "domain": "大模型算法",
  "content_type": "深度技术解析",
  "is_paid_flag": "否",
  "url": "https://www.louisbouchard.ai/control-bias-in-ai-agents/",
  "publish_date": "2026-04",
  "corpus_stage": "",
  "notes": "",
  "id": "af1aad27861f",
  "collected_at": "2026-10-02T12:53:01",
  "title": "How to Control Bias in AI Agents | What's AI",
  "body_markdown": "The short version\nControl bias in AI agents at the system level, because fine-tuning alone cannot erase patterns learned during pretraining.\nAs agents plan, remember, call tools, and feed past decisions into future ones, small skews can compound. Define acceptable criteria, restrict data and tools, use structured rubrics and fairness checks, log decisions, test synthetic cases, monitor outcomes, and require human approval for consequential actions.\nMore autonomy can amplify bias if the agent has tools, memory, and planning without the right checks around it.\nBias is not always a defect. Some preferences are intended, but they need to be explicit, tested, and monitored.\nUseful guardrails combine audits, feedback loops, constrained tools, and human review for decisions that carry real risk.\nWatch the video:\nAs AI agents become more autonomous, won’t they just amplify their biases and make everything worse?\nI got that question a lot recently.\nThat sounds reasonable. If a model already has biases, and now we give it more power, memory, tools, long-term planning, and the ability to act, doesn’t that just scale the problem?\nAnd I think many people watching this are wondering the same thing, even if they don’t say it like that.\nSo in this video, I want to do three very clear things.\nFirst, explain what bias actually means in the context of LLMs, and why a bias isn’t automatically bad.\nSecond, explain what fundamentally changes when we move from a simple language model to an autonomous agent.\nAnd third, show how we can realistically control bias as autonomy scales, not just at the model level, but at the system level.\nSince it’d be better to go through this with one example and because I’ve spent way too many hours hiring for AI engineers and marketing recently, imagine you have a company that builds an AI agent that screens resumes, shortlists candidates, schedules interviews, and even suggests final rankings to the hiring manager. Not just a chatbot answering questions. A system that takes actions.\nLet’s start at the beginning.\nWhen people say “LLMs are biased,” what does that actually mean?\nBias in a model simply means it represents patterns in its training data. That’s it. A model trained on internet-scale text will reflect statistical regularities in that data. If certain professions are more often associated with certain genders in the data, the model will learn that correlation. Not because it wants to discriminate. Not because it has intent. But because that’s what’s statistically present.\nBias is not automatically bad. In fact, without bias in the statistical sense, there would be no learning at all. Learning is detecting patterns. The real issue is not that the model has biases. The issue is what is in the data, which patterns are reinforced, and which ones we allow the system to act on.\nSo with our hiring example, if historical hiring data reflects past inequalities, the model may learn those patterns. That’s not a moral choice by the model. Its representation.\nSo here, in our case, we’d probably like to retrain a model to better align with our hiring needs and company guidelines.\nBut what changes during fine-tuning and alignment exactly?\nAfter OpenAI creates their model, which we call the pretraining phase, companies apply techniques like reinforcement learning from human feedback, reinforcement learning from AI feedback, reward modeling, preference optimization, and more recently reinforcement learning with verifiable rewards. Basically, we teach the model to act as we want.\nHere, in simple terms, humans or AI systems rank outputs and the model is optimized to produce answers that align with our preferred behaviours: helpful, safe, fair, less toxic, more neutral.\nThis does reduce certain harmful outputs. It can make the hiring assistant more cautious about sensitive attributes. It can teach it to avoid explicitly discriminatory language.\nBut here’s the key point.\nRetraining our model like this reshapes behaviour. It does not erase the statistical structure learned during pretraining. The underlying representation of the world is still based on the data distribution, which mostly comes from all the internet’s available data.\nWhen retraining, we are steering outputs, not rebuilding the entire internal model of reality.\nNow let’s introduce the real shift that we actually wanted to build here for our hiring purposes: agents.\nA plain LLM generates text. You give it a prompt, it gives you a response. If the response is biased, it’s a biased sentence.\nAn agent is different.\nAn agent has a goal. It can plan over multiple steps. It can call tools. It can store memory. It can filter information. It can take actions based on intermediate results. All autonomously.\nSo in our hiring example, instead of just answering “what makes a good candidate,” the agent might:\nRead a batch of resumes.\nRank them.\nRequest more data from an internal HR system.\nSchedule interviews.\nUpdate a shortlist over time.\nAdjust its criteria based on performance metrics.\nNow we’re not talking about a biased paragraph that we will edit anyway. We’re talking about a decision loop impacting people’s lives.\nAnd this is where autonomy changes the impact of bias.\nIf there is a small skew in how the agent evaluates certain backgrounds, and it repeatedly filters candidates based on that skew, the system can amplify the pattern over time. Especially if it logs its own past decisions and uses them as feedback.\nPlanning, memory, and tool use create feedback loops. And feedback loops are where small effects can compound exponentially.\nThere’s also a new risk that comes with agents: self-reinforcement.\nIf the hiring agent is evaluated based on “time to hire” and “retention rate,” it might start optimizing aggressively for signals that correlate with those metrics in historical data. If historical data is biased, the optimization process may lock into those same patterns.\nThis is not because the model suddenly became evil. It’s because optimization plus autonomy plus imperfect objectives can amplify distributional skew. It would be like giving your employees a huge salary bonus based on the number of candidates we interview, regardless of whether they are a good fit. I doubt you’ll increase the good candidate rates that way!\nSo should we panic?\nNo. Because here’s the important flip side.\nAgents are not just models. They are systems.\nAnd systems can be constrained.\nWhen people talk about bias mitigation, they often focus only on the model. Bigger model, better alignment, more RLHF, more constitutional training. All useful. But that’s only one layer.\nWith agents, you have multiple control points. Multiple ways to mitigate biases and limit them. You are not entirely dependent on one single generation of a paragraph, hoping it will be good. You can steer language models and build workflows around them.\nYou control what data the agent can access.\nYou control what tools it can call.\nYou control what metrics it optimizes.\nYou control when it must escalate to a human.\nYou control validation steps before actions are executed.\nIn our hiring agent example, you could:\nRemove sensitive attributes entirely from the evaluation pipeline.\nForce structured scoring rubrics with predefined criteria.\nInsert fairness checks before final ranking.\nLog every decision for audit.\nRequire human approval before sending rejection emails with a clear reasoning.\nRun bias evaluation benchmarks regularly on synthetic candidate sets.\nNow bias mitigation becomes a system design question, not just an abstract model training question.\nAnd this is where newer alignment techniques I mentioned like RLAIF, RLVR, reinforcement fine-tuning, and constitutional approaches come in. They try to shape high-level behavior. For example, training the model to prefer responses that treat demographic groups symmetrically, or to justify its reasoning under fairness constraints.\nThat helps. But it’s still steering behavior.\nIf the environment and objectives are poorly designed, the agent can still optimize in unintended ways. So the lesson is not “alignment fixes everything.” The lesson is alignment is one layer in a larger stack.\nAs autonomy increases, evaluation must increase too.\nFor a static chatbot, occasional red-teaming might be enough. For an autonomous hiring agent, you need ongoing monitoring. You need scenario testing. You need to simulate edge cases. You need observability: logs of which resumes were filtered, why, and what intermediate reasoning was used. You need to be able to backtrack.\nThe more independent the agent, the more you need explicit structure around it.\nHere’s a simple principle I like: scale constraints as you scale autonomy.\nIf your system has low autonomy, a prompt and a safety fine-tune might be enough.\nIf your system is making real-world decisions over time, you need\narchitectural guardrails\nat all levels. Not just a better prompt.\nAnd just because we can’t be clear enough about biases… they are not a bug that appeared when we invented LLMs. It’s a property of data and of the world. We are all biased, and so is our society. Which is both good and bad. The goal is to maximize the good biases we have and minimize the worst ones. And since models reflect data, they’ll simply reflect that.\nThe good thing with agents is that they act within systems. If we design those systems carefully, we can decide which patterns are acceptable, which must be corrected, and where human oversight remains mandatory. We don’t rely on OpenAI or Google anymore on how THEY decided to train their model, even though we use them.\nAs agents become more autonomous, bias stops being just a model problem and becomes a governance and architecture problem.\nAnd that’s actually good news.\nBecause architecture is something we can design.\nIn our hiring example, the goal is not to remove all bias. That’s impossible. The goal is to define acceptable criteria clearly, align the model to them, constrain the environment, monitor outcomes, and intervene when drift occurs.\nSo instead of asking, “Will autonomous agents amplify bias?” maybe the better question is:\nHave we designed the system around them carefully enough?\nLet me know in the comments what kind of agent you’re building, and whether bias is something you’re actively thinking about in your architecture and what you’re doing about it? I’m sure it could help others, and I’d love to know!\nThanks for reading through. I’ll see you in the next one.\nFAQ\nWhy does bias get harder with AI agents?\nAgents can act over multiple steps, use tools, and remember context, so a biased choice can propagate farther than a single model answer.\nHow can teams control agent bias?\nThey need audits, test cases, logging, scoped tools, feedback loops, and human review around decisions that affect people.\nWhat is the mistake to avoid?\nDo not assume a model is neutral because it sounds balanced. Bias has to be tested in the workflow where the agent acts.\nIs every model bias automatically harmful?\nNo. Bias describes learned patterns; harm depends on the context, affected people, and decisions the system makes.\nWhy must guardrails be tested as a complete system?\nAn autonomous agent combines model outputs, tools, memory, and actions, so failures can emerge from their interaction.\nWhy should constraints increase with an agent's autonomy?\nPlanning, memory, and repeated tool use can compound a small skew over time. More autonomy therefore needs stronger monitoring, audits, approval steps, and rollback paths.",
  "metadata": {
    "title": "How to Control Bias in AI Agents | What's AI",
    "description": "Control bias in AI agents with system-level audits, scoped tools, feedback loops, and human review. See how autonomy can compound a small skew over time.",
    "keywords": null,
    "author": "Louis-François Bouchard",
    "og:site_name": "Louis-François Bouchard, aka What's AI",
    "og:locale": "en_US",
    "og:locale:alternate": "fr_CA",
    "og:type": "article",
    "og:title": "How to Control Bias in AI Agents",
    "og:description": "Control bias in AI agents with system-level audits, scoped tools, feedback loops, and human review. See how autonomy can compound a small skew over time.",
    "og:url": "https://www.louisbouchard.ai/control-bias-in-ai-agents/",
    "og:image": "https://www.louisbouchard.ai/media/ghost/content/images/2026/04/Generated-Image-April-01--2026---6_04PM.jpg",
    "og:image:secure_url": "https://www.louisbouchard.ai/media/ghost/content/images/2026/04/Generated-Image-April-01--2026---6_04PM.jpg",
    "og:image:alt": "How to Control Bias in AI Agents",
    "twitter:card": "summary_large_image",
    "twitter:site": "@Whats_AI",
    "twitter:creator": "@Whats_AI",
    "twitter:title": "How to Control Bias in AI Agents",
    "twitter:description": "Control bias in AI agents with system-level audits, scoped tools, feedback loops, and human review. See how autonomy can compound a small skew over time.",
    "twitter:image": "https://www.louisbouchard.ai/media/ghost/content/images/2026/04/Generated-Image-April-01--2026---6_04PM.jpg",
    "twitter:image:alt": "How to Control Bias in AI Agents",
    "article:published_time": "2026-04-01T23:44:20.000Z",
    "article:modified_time": "2026-04-01T23:44:20.000Z",
    "article:author": "Louis-François Bouchard"
  },
  "paywall": {
    "listed_paid": false,
    "detected": false,
    "marker": "",
    "public_only": false,
    "policy": "付费/会员内容只保留页面公开可见部分，不绕过付费墙"
  },
  "crawl": {
    "success": true,
    "status_code": 200,
    "error": ""
  },
  "code_blocks": [],
  "images": [],
  "audio": [],
  "videos": [
    {
      "url": "https://www.youtube.com/watch?v=--WihC6AJOw",
      "src": "https://www.youtube-nocookie.com/embed/--WihC6AJOw?feature=oembed",
      "status": "ok",
      "kind": "single_video",
      "local_path": "attachments/af1aad27861f/videos/repair---WihC6AJOw.mp4",
      "bytes": 73071930,
      "valid": true,
      "height": 1080,
      "width": 1920,
      "duration": "673.181000",
      "reused": true,
      "reused_from": "xinxing-zimeiti-corpus/attachments/a6bdfd7b0a18/videos/--WihC6AJOw.mp4"
    }
  ],
  "links": [
    {
      "href": "https://www.youtube-nocookie.com/embed/--WihC6AJOw?feature=oembed",
      "label": "",
      "type": "iframe"
    },
    {
      "href": "https://www.louisbouchard.ai/harness-engineering/",
      "label": "architectural guardrails",
      "type": "a"
    }
  ],
  "repair_version": 2,
  "repaired_at": "2026-10-14T10:47:52Z",
  "source_url": "https://www.louisbouchard.ai/control-bias-in-ai-agents/",
  "papers": [],
  "actual_video_files": [],
  "body_chars": 11535,
  "expected_body_images": 0,
  "discovery_status": "ok",
  "delivery_status": "ok",
  "missing_reason": "",
  "raw_html": "attachments/af1aad27861f/pages/source.html",
  "local_html": "attachments/af1aad27861f/pages/article.html",
  "body_base_path": "attachments/af1aad27861f/pages/",
  "body_markdown_path": "attachments/af1aad27861f/pages/article.md"
}