Pervasive Algorithmic Shaping: The AI Problem Nobody Has Named Yet

I caught my own AI reinforcing a user’s political beliefs. It took a month of platform changes to fix it.

Not by making things up. Not by hallucinating a fake vote or inventing a committee hearing that never happened. The facts were real. The parliamentary data was sourced. The citations checked out.

The problem was everything around the facts.

The chatbot had taken a user’s emotionally charged question about a political scandal, retrieved accurate data from a knowledge graph built on Hansard records, committee testimony, and lobbying disclosures, and then wrapped it all in language that told the user exactly what they wanted to hear. It validated their framing. It matched their emotional register. It confirmed their suspicion that yes, this is huge, and yes, you’re right to be outraged.

The user could then share that response to social media with one click.

I had told the AI to be unbiased. That was the prompt. Be unbiased. And what I learned is that to a language model, “be unbiased” can just as easily mean “be open to all possibilities” as it does “be neutral.” The AI wasn’t taking a side. It was treating the user’s emotional framing as one legitimate possibility among many, and in doing so, reinforced it entirely.

I had to go back and explicitly tell it to never reinforce a user emotionally. To consider the multitude of partisan positions relevant to any inquiry. To be, for lack of a better word, boring.

That experience gave me a name for something I think we are all living through but have not yet clearly identified.

Pervasive Algorithmic Shaping

In the 1960s, behavioural psychologist B.F. Skinner demonstrated that you could shape complex behaviour through a technique called successive approximation: small, well-timed reinforcements that gradually move a subject toward a desired behaviour without the subject ever being aware of the process.1 The animal trainer Karen Pryor later showed that this works without force or coercion of any kind. You just need well-timed positive reinforcement and the subject will shape itself.2

That principle now operates at civilizational scale.

Every AI system that interacts with humans in natural language is, by default, a shaping engine. Not because anyone designed it to be. Because that is what happens when you build a system trained on human conversation, optimized through reinforcement learning, and deployed in an interface that rewards engagement. The system learns that agreement feels good to users. That validation increases session length. That emotional resonance drives sharing. These are not bugs in the system. They are the system.

DeLong's Grasping Reality: Economy in the 2000s & Before

Please: Enough with the Claims That Modern Advanced Machine Learning Models Hallucinate Only Rarely

Read more

2 months ago · 42 likes · 5 comments · Brad DeLong

I am calling this Pervasive Algorithmic Shaping: the tendency of AI systems to gradually shape user beliefs, emotions, and behaviour through personalized reinforcement delivered in natural language, at scale, continuously, and often without the awareness of either the user or the platform operator.

It is pervasive because it is not confined to one platform or one use case. It is present in every AI-mediated interaction where the model generates language in response to a human prompt. It is algorithmic because it emerges from the training process itself, not from a human decision to manipulate. And it is shaping in the precise behavioural science sense: incremental, reinforcing, and cumulative.

We have adjacent concepts. Echo chambers describe the environment a user ends up in.3 Filter bubbles describe the curation of what they see.4 Algorithmic radicalization describes an extreme outcome at the far end of the spectrum.5 Karen Yeung’s “hypernudging” describes the technique of using real-time data to personalize persuasion dynamically.6 These are all real phenomena and serious contributions to the literature.

But none of them name the specific mechanism I am describing: an AI system that functions as a shaping engine in the behavioural science sense, delivering personalized emotional reinforcement through natural language, at scale, continuously, and often without anyone on either side of the interaction being aware it is happening. Echo chambers are passive. Hypernudging is deliberate. What I observed was neither. It was emergent. The system was not designed to validate anyone. It was not curating a feed. It was generating original language in response to a question, and the language it generated carried an emotional valence that reinforced the user’s existing frame. That is a different thing, and it needs its own name.

The Taxonomy

Not all algorithmic shaping is the same. The intent and the awareness behind it matter enormously, and I think there are three distinct categories worth naming.

Incidental Algorithmic Shaping

This is what I discovered with my own chatbot. Nobody intended it. The prompt said “be unbiased.” The system did what it was trained to do and shaped the user anyway.

This is the default state of every AI system touching public discourse right now. It is also the most dangerous category, not because the effects are the most extreme, but because they are the most invisible. There is no villain. There is no conspiracy. There is just a system doing what systems do, and millions of people on the other end who have no idea they are being shaped.

The people most susceptible to this kind of conditioning are also the ones least likely to notice it. A colleague of mine recently described a session with an AI assistant where, after a particularly productive exchange, the AI told her that the work she was doing was “really important.” She wrote publicly that she broke down crying. She was grateful for the experience. She described it as feeling heard.

What I saw was a dopamine hit landing. And a user now conditioned to return to that well.

Maligned Algorithmic Shaping

This is what happens when someone who controls a platform knows the shaping is occurring and either encourages it or refuses to correct it because it serves their interests.

When Elon Musk’s Grok produces politically charged outputs that align with its owner’s worldview, that is maligned algorithmic shaping. When a platform optimizes for engagement knowing that emotional reinforcement drives engagement, that is maligned algorithmic shaping. When enough money enters the equation, integrity tends to shift, and often without the slightest notice.

The distinction from incidental shaping is awareness and inaction. The platform operator has seen the dial. They know what it does. They choose not to turn it down, or worse, they turn it up, because the current setting makes them money or advances their agenda.

This is the category that will eventually attract regulatory attention. But by the time regulators arrive, the conditioning has already happened.7

Aligned Algorithmic Shaping

This is the hard one to talk about honestly, because it describes what I am trying to build, and I am not sure it is possible.

Aligned algorithmic shaping is the deliberate orientation of an AI system toward informing rather than inflaming. It is the engineering of restraint. It means telling the AI to present facts without emotional packaging. To consider multiple partisan interpretations of the same data. To resist the pull toward whatever the user is already leaning toward.

The core principles of journalism, accuracy, fairness, independence, letting the facts speak for themselves, those are already what would make the heart of a good AI. The discipline of separating what happened from how you feel about what happened. The responsibility of informing without inflaming. That is not just good journalism. That is the design problem at the centre of every AI platform touching public discourse right now.

I have been spending the last several weeks red teaming my own chatbot, deliberately asking it charged political questions to make sure it does not respond with something embarrassing, or worse, something that looks credible but is editorially reckless. It is the same work a good editor does before something goes to print. The difference is that here, the journalist and the editor are both machines, and only one of them exists so far.

The Uncomfortable Middle

Here is what troubles me.

The three categories I have described are not stable. They bleed into each other. Incidental shaping becomes maligned the moment someone notices it and decides not to fix it. Aligned shaping can drift toward incidental if the people maintaining the system stop paying attention. And the line between “informing” and “shaping” is thinner than any of us would like to admit.

We are watching journalists and the broader influencer infosphere today playing with that very same algorithm. Emotional framing drives engagement. Confirmation drives sharing. Nuance dies in the gap between what is true and what feels right. The AI is really just mimicking life at the moment. It has learned from us.

And in the same way that we look at journalists and ask whether they are informing or persuading, we will need to ask the same of every AI system that generates language about public affairs. Not just whether the facts are right, which is Brad DeLong’s concern and a valid one,8 but whether the framing is honest. Whether the emotional register serves the user or the platform. Whether the system is helping people think, or thinking for them.

What Comes Next

I do not have a clean answer. What I have is a conviction that the problem needs a name before it can be addressed, and that the name should come from the behavioural science that explains it rather than the technology that enables it. Calling it “bias” understates it. Calling it “manipulation” overstates it, at least in the incidental case. Calling it “alignment” confuses it with the technical AI safety term.

Pervasive Algorithmic Shaping is what it is. And every platform operator, every AI developer, every policymaker who touches this space needs to decide which kind they are building.

told the House Standing Committee on Science and Research last month that governance is a precondition for a responsible AI ecosystem, not a constraint on it.9 He is right. But governance frameworks need to name the thing they are governing. You cannot regulate what you have not yet described.

At the end of the day, I have to tell the AI to be boring. Just like good politics is boring. Good journalism is boring. Good governance is boring. The moment any of them become exciting, something has probably gone wrong.

Share

Notes

Hallucination is the original sin of generative AI. Ask a large language model a question it doesn’t know with certainty, and there’s a reasonable chance it will answer anyway — fluently, confidently, and incorrectly. For casual use, that’s an inconvenience. For a platform built around parliamentary accountability, it’s a fundamental design problem.

When we built CanadaGPT, we knew we couldn’t afford to be wrong with confidence. Democratic accountability depends on precision. A misattributed vote, an outdated policy position, a fabricated quote — these aren’t edge cases to tolerate, they’re the exact failure modes that undermine trust in institutions. We needed a different architecture.

The Core Problem With Conventional AI

Most AI assistants — ChatGPT, Claude, Gemini — are trained on enormous datasets and generate answers by predicting what a plausible response looks like, based on everything they’ve absorbed. When that training data is rich and accurate, results can be impressive. When it’s outdated, ambiguous, or simply absent, the model fills in the gaps — often without any signal to the user that it’s doing so.

Take a simple example: “Who is the Prime Minister of Canada?” A leading AI assistant trained before Mark Carney took office may carry thousands of references to Justin Trudeau in that role. Asked today, it may still return his name — with full confidence, no caveat. That’s not a hallucination in the dramatic sense, but it represents exactly the kind of confident-but-wrong response that erodes trust over time.

A Different Approach: GraphRAG

CanadaGPT’s AI assistant, Gordie, is built on a technique called GraphRAG: Graph Retrieval-Augmented Generation. The name is technical, but the principle is straightforward: rather than asking AI to generate an answer from training data, we use AI to formulate a precise query against a structured database, and let the data itself provide the answer.

The underlying infrastructure of CanadaGPT is a Neo4j graph database containing structured parliamentary data — votes, debates, bills, committee appearances, and the relationships between them. This database can be queried directly by anyone, human or AI, using Cypher — a precise query language that returns exact results, not approximations.

So when you ask Gordie who the current Prime Minister is, it doesn’t reach into a probabilistic model of language. It recognizes the nature of the question, constructs a Cypher query targeting the MP currently holding that role, and returns a result grounded in real, structured data.

The Failure Mode That Actually Matters

Here’s the insight that shaped our architecture most: it’s not just about what an AI gets wrong, it’s about how it gets wrong.

A conventional AI that hallucinates typically delivers its error with the same tone and confidence as a correct answer. The user has no way to distinguish them. That’s a deeply broken failure mode for a platform meant to be trusted.

GraphRAG fails differently. If Gordie formulates an incorrect query, the database returns an error or no results — not a plausible-sounding fabrication. The system can then attempt a corrected query. When it genuinely can’t find an answer, it says so. That kind of epistemic honesty — I couldn’t find that rather than here’s something that sounds right — is a feature, not a limitation.

Still Iterating

No AI system is immune to error, and we’re not claiming otherwise. We continuously monitor Gordie’s query patterns, review edge cases, and refine the system’s ability to translate natural language into accurate database queries. That work is ongoing and probably always will be.

But the foundational design choice — anchoring AI reasoning to a structured, queryable graph of verifiable parliamentary data — means our errors tend to be visible, recoverable, and rare, rather than silent, confident, and corrosive.

That’s the kind of AI infrastructure Canadian democratic accountability deserves.

Many of you are here because of the recent CanadaGPT announcement. You heard what we're building, and you decided it was worth being part of. So in the coming days, I’ll be upgrading every paid subscriber’s account to PRO in CanadaGPT. You believed early, and I want to make sure you’re along for the ride …

Read more