AI Inherits Human "Addition Bias"

Researchers at the University of Tübingen just documented something everyone using ChatGPT has probably noticed but couldn't quite articulate: AI language models have a systematic tendency to make problems more complicated by adding information, even when removing information would be simpler and more effective.

This isn't a quirk. It's a cognitive bias called "addition bias"—the human tendency to solve problems by adding elements rather than subtracting them, even when subtraction is more efficient. And according to research published in Communications Psychology, large language models (LLMs) like GPT-4 and GPT-4o don't just exhibit this bias—they amplify it beyond human levels.

The Research: Testing Humans Against AI

Lydia Uhler, Verena Jordan, and colleagues ran two studies comparing human responses to LLM outputs across spatial and linguistic tasks. Study 1 involved 588 human participants versus 680 GPT-4 responses. Study 2 compared 751 humans to 1,080 GPT-4o outputs.

The tasks were designed so that in some trials, adding information solved problems more efficiently, while in others, removing information was clearly better. Instructions were written using either neutral or positive language to test whether framing affected bias.

Spatial tasks required arranging shapes or structures in specific ways. Linguistic tasks involved choosing or generating text following instructions, like improving an essay by either adding explanations or removing unnecessary sections.

The results were consistent: both humans and LLMs showed addition bias. But LLMs exhibited the bias more strongly, particularly on tasks where subtraction was objectively more efficient.

When Subtraction Is Better, AI Adds More Anyway

The critical finding: humans made fewer additive choices when subtraction was clearly more efficient than addition. They adapted their strategy based on task demands. GPT-4 showed the opposite pattern—it increased additive responses even when subtraction was demonstrably better.

GPT-4o performed slightly differently. On linguistic tasks, it aligned with human patterns. On spatial tasks, it showed no efficiency effect whatsoever—meaning it didn't adjust strategy based on whether addition or subtraction was more effective. It just defaulted to addition regardless.

Instruction framing also mattered. When instructions used positive valence (encouraging, optimistic language) rather than neutral language, both GPT models generated more additive outputs in linguistic tasks. Humans showed this pattern only in Study 2. The models were more susceptible to linguistic framing than humans.

This suggests LLMs aren't just mirroring human biases—they're amplifying them and applying them more rigidly than humans do.

Why This Happens: Training Data Inheritance

LLMs are trained on massive corpora of human-written text. If human writing exhibits addition bias—favoring elaboration, additional context, and more explanation over concise editing—then models trained on that data will inherit and reproduce those patterns.

But the amplification is the concerning part. Humans demonstrated some ability to recognize when subtraction was more efficient and adjust accordingly. GPT-4 did the opposite, becoming more additive precisely when subtraction would have worked better. This suggests the training process doesn't just capture human biases—it may even reinforce them by optimizing for certain types of responses that feel more complete or helpful.

Consider typical ChatGPT usage: you ask for help improving a document. The model's default response is almost always to add more context, more examples, more explanation, more qualification. Rarely does it suggest removing sections, even when documents are clearly bloated with unnecessary information.

This aligns perfectly with the research findings. The model has learned that human-written "improvements" typically involve addition, so it defaults to that strategy even when subtraction would be more effective.

The Practical Implications

This matters beyond academic curiosity. According to the Wharton cognitive surrender research published earlier this month, people adopt AI outputs without sufficient scrutiny roughly 80% of the time. If AI systematically overcomplicates problems through addition bias, and humans uncritically adopt those overly complex solutions, we're not augmenting human judgment—we're making decisions more complex than necessary.

In business contexts:

• AI-generated reports become longer and less focused because the model defaults to adding sections rather than cutting unnecessary ones
• Problem-solving recommendations become more elaborate when simpler interventions would be more effective
• Strategic planning gets bloated with additional considerations rather than being focused on core priorities

The addition bias also compounds with the confidence problem. AI delivers these overcomplicated solutions fluently and confidently. Users don't question whether simpler approaches might work better—they accept the complex solution because it came from the AI and sounds authoritative.

What the Researchers Recommend

Uhler, Jordan, and colleagues emphasize that this finding should guide the development of more reliable AI agents. Understanding that LLMs inherit and amplify human cognitive biases means:

• Training approaches need to explicitly counter addition bias, not just optimize for human-like output
• Models should be evaluated on whether they choose appropriately between addition and subtraction strategies based on task demands
• Interface design should prompt users to consider whether AI suggestions are overcomplicated

The researchers also note that this work should inform a better understanding of human decision-making patterns. If AI makes addition bias visible and measurable at scale, we can study it more systematically and develop interventions.

The Pattern Across Multiple Studies

This research converges with other recent findings about AI reasoning patterns:

Stanford's productivity research showed AI enables output increases, but didn't examine whether that output was appropriately complex or systematically over-elaborated. The Wharton cognitive surrender study found that humans don't sufficiently scrutinize AI outputs. Now the Tübingen addition bias research shows AI systematically makes problems more complex than necessary.

Put these together: AI generates overcomplicated solutions, delivers them confidently, and humans adopt them without questioning whether simpler approaches would work better. That's not augmentation—it's systematic bias amplification with cognitive surrender preventing correction.

What Users Should Actually Do

If you're using LLMs for problem-solving, decision-making, or content generation:

1. Explicitly ask for subtraction: "What could I remove to make this simpler?" rather than just "How can I improve this?"
2. Question additions: When AI suggests adding elements, ask whether removing existing elements might be more effective
3. Test simpler versions: Before implementing AI's recommendations, try the simplest possible version first
4. Recognize the bias: Understanding that models default to addition helps you compensate when evaluating outputs

The goal isn't to stop using AI—it's to use it with awareness that it inherits human biases and sometimes amplifies them. Cognitive surrender becomes less likely when you know what patterns to watch for.

AI implementation requires understanding systematic biases in model outputs, not just capability claims. Winsome Marketing's growth experts help you evaluate AI tools through the lens of actual decision-making patterns, not vendor promises about intelligence. Let's talk about AI strategies that account for cognitive biases in both humans and models.