Steering

An AI system that writes text has a certain behavior after its learning process. It answers in a particular tone, sometimes elaborate, sometimes terse, sometimes cautious. Steering is the umbrella term for all techniques used to shift this behavior afterward in a desired direction. The English word simply means to guide, as with the helm of a ship. The comparison fits well: you don’t rebuild the ship, you turn the rudder. In narrower technical usage, however, Steering means something specific, namely direct intervention in the numerical values that arise inside the system while it generates a response.

Steering instead of retraining

Training a large AI model from scratch costs millions and takes weeks. If a provider notices that its system responds too submissively or dodges too often, it cannot start over every time. Steering is the inexpensive way: you change the behavior without touching the model itself. Some methods even work instantly and can be switched off again at will.

The second reason is safety. A model should not provide instructions for weapons and should not insult people. Such boundaries are already set during training, but they don’t always hold. Steering techniques offer an additional layer that counteracts when needed. Research groups at Anthropic and OpenAI regularly publish work on this, because it allows both safety and cost to be improved at the same time.

It is important to distinguish this from finetuning. In finetuning, the model is further trained with additional examples and permanently changed. Steering in the narrower sense leaves the model unchanged and intervenes only at the moment of the response. Both pursue the same goal, but differ in effort and reversibility.

Reaching into the intermediate layers

While a model generates a response, the information travels through many computational stages. At each stage, long sequences of numbers arise, called activations. Researchers have discovered that certain properties can be found within these numbers. For example, there is a typical numerical direction for polite phrasing and another for reticence on sensitive topics.

These directions are found by presenting the model with many example pairs. Once with polite sentences, once with impolite ones. The difference between the activations yields what is called a steering vector. This vector is then added to the running numerical values during the response. The model then writes more politely, without anyone having typed an instruction to do so.

A common misconception is to mistake this for precise remote control. In reality, the effect is crude. If you turn the strength up too high, the model becomes obsessed with its topic and produces garbled output. A well-known experiment by Anthropic amplified the concept of the Golden Gate Bridge so strongly that the model worked the bridge into almost every response. Simpler methods also count as steering, such as fixed system instructions running in the background of a chatbot.

Steering in chatbots and headlines

Every chatbot you use is already being steered. Before your question, there is an invisible text that gives the model its role, tone, and prohibitions. If an assistant stubbornly refuses to make medical diagnoses, exactly such an instruction is often behind it. Developer tools also offer sliders for creativity or verbosity.

In the news, the term usually comes up in connection with AI safety. When a provider announces that its model is now less sycophantic, steering is often behind it. Conversely, security researchers report on attacks that abuse the same techniques to disable safety mechanisms. Both sides work with the same tools.

For investors and companies, steering is above all a matter of cost. Anyone wanting to adapt a model for their own customer service does not necessarily need to retrain it expensively. Often, good instructions and targeted interventions are enough. This significantly lowers the barrier to entry for smaller companies.