What is steerable AI, why does it matter and, how do you build it?

Adjusting AI systems and explaining their outputs is not always easy—make sure your strategy accounts for 'steerability' needs.
MJ Petroni
July 12, 2023

AI is exciting—and also sometimes hilariously bad, or even frightening! Steerable AI refers to the idea of being able to fine-tune, adjust, correct, or otherwise guide an AI system to operate more in line with the expectations and ethics of its owner or user.

In some AIs, once the underlying model is trained and deployed, the AI operates autonomously without much user intervention, like the systems in basic motion-detecting lights. These systems might be comparatively static (in other words, they don't learn once released) or self-training (by observing user behaviors or other inputs—like an automatic car transmission that 'learns' its drivers' habits).

Steerable AI, on the other hand, aims to incorporate explicit user preferences, ethical considerations, or business requirements into the decision-making process of the AI system. Steerable AI systems often employ techniques such as rule-based systems (sometimes known as expert systems), so-called human-in-the-loop feedback, or interactive interfaces to modify or guide the behavior of the AI system. For example, recommendation engines like Netflix's movie suggestions allow users to explicitly select their favorite 'seed' movies and then up-rate or down-rate later suggestions.

Hands holding a mobile phone displaying Netflix recommendations.
Netflix' recommendations are user-steerable based on initial account setup and adjustments on the user's settings page.

These techniques enable users to adjust the AI's actions, policies, or outputs based on changing circumstances or evolving needs. The concept of steerable AI is particularly relevant where human input is desired, such as autonomous vehicles, health/medical diagnoses, or algorithmic decision-making in sensitive domains like finance or law. Steerable AI strategies provide a mechanism to align the AI system's behavior with human values, ensuring transparency, accountability, and adaptability.

What kinds of AI can be steered?

Traditional analytical algorithms are generally easier to steer—and explain—than machine learning and generative AIs, which have many parameters that can be difficult to make accessible.

An expert system refers to a set of 'if this, then that'-style instructions, like a decision tree, that help a machine respond to data passed to it by other systems or humans. When your thermostat turns on the air conditioning in response the temperature rising, that's a rule, and a number of them together can result in something like the classic WebMD symptom checker—which asks users a series of questions and then suggests potential ailments and remedies in response, refining the results with further questions. These expert systems are sometimes called AI in marketing, even though there's nothing inherently intelligent about them. They are doing a 'dumb' recall of a smart human expert's knowledge for the user. These systems are able to be explained and steered quite easily, because the entire logic of the system can be reverse-engineered quite easily.

Analytical tools, like the technologies used to analyze users' web traffic and then provide them with basic recommendations about other sites, stories or products they might like to see, are often just a very complex expert system as we described above. They too can be steered fairly easily.

However, due to their complexity, more advanced machine learning and deep learning approaches become harder and harder to explain and steer. For example, unsupervised machine learning finds patterns in data without being guided to do so by a human and could find unexpected correlations in a set of photos akin to 'people who wear polka dots eat oatmeal' that don't really make sense until carefully assessing the training data and finding out that a lot of the source images which had polka dots were from one photo shoot for an oatmeal advertisement. Other times, the correlations are completely wrong and need to be corrected. Deep learning, which refers to many layers of machine learning, is exponentially more complex and thus sometimes quite hard to decipher and steer.

Generative AI, which is based on massively complex models of human language and other data, like images, creates outputs that can't always be understood. Steering generative AI is often done by carefully selecting the underlying training data (like which websites' content to import or using a company's specific product information) and then, afterward, with human reinforcement learning. The strategies for steering generative AI are worth an article of their own, so we won't get into them fully here.

How do you build steerable AI?

Here are some steerable AI strategies, roughly in order:

Define goals and constraints

Determine the objectives, goals, and constraints the AI system should adhere to. These could be ethical guidelines, legal requirements, user preferences, or specific performance metrics. This involves what you can think of as computational thinking—being able to decompose or deconstruct a problem, create an abstraction or model of a system, identify reusable patterns, and then recombine these elements into working algorithms and programs.

Collect and annotate data

Gather relevant data that represents the problem domain and the desired outcomes. For example, you might pre-label a large number of images and annotate the data with labels or additional information that can guide the AI system's behavior. This annotated data will be used for training and evaluation. (You can read more about data's lifecycle in the Data Supply Chain Guidebook.)

Train models

Train an initial AI model using the annotated data. The choice of training data, machine learning algorithms, and techniques depends on the specific problem and available resources. (For examples of data sources that could be used for training, check out the Data Sources Explorer). Some deep learning methods, such as neural networks, are commonly used due to their ability to learn complex patterns—but may also be harder to adjust. (Read more in our Guidebook to Machines That Learn).

Set and adjust rules

Integrate rule-based systems or decision logic to steer the AI system's behavior. These rules can be predefined based on domain knowledge or specified by users to ensure specific actions or constraints are enforced.

Incorporate feedback

Enable user interaction and feedback loops to gather input from users or domain experts. This feedback can fine-tune or modify the AI system's behavior. Techniques like active learning or reinforcement learning can be employed to optimize the model based on your preferences.

Create interfaces

Generative AI systems are often based on text prompts, like "tell me a story," and the user interface can include micro-training for users on how to write better prompts with more detail or more precision, and advanced systems might also directly coach the user on how to better ask questions of the AI system. But text-based prompts are hard for everyday users to organize and store for consistent use, and many tools, like ChatGPT, currently do not have "prompt libraries" or "prompt chain builders" built-in. There's a big opportunity to help users get more consistent results with these interfaces.

Prompt-chaining: sequencing a number of prompts for an AI tool, like a generative AI chat system, in order to accomplish a given goal.

User interface is also a prime location to raise user's AI and Digital Fluency. For example, OpenAI's ChatGPT integrates cautions about the limits of generative AI into its user responses when they ask for facts or information on current events. Similarly, good user experiences help train models and provide feedback to AI developers, as with ChatGPT's 'thumbs-up, thumbs-down' interface next to every response.

Screenshot of ChatGPT opening screen. Three columns: 1) Examples: "Explain quantum computing in simple terms", "Got any creative ideas for a 10 year old’s birthday?", "How do I make an HTTP request in Javascript?"; 2) Capabilities: Remembers what user said earlier in the conversation, Allows user to provide follow-up corrections, Trained to decline inappropriate requests; 3) Limitations: May occasionally generate incorrect information, May occasionally produce harmful instructions or biased content, Limited knowledge of world and events after 2021.
ChatGPT's opening screen for a new chat shows examples of prompts, capabilities of the tool and limitations of the system—building AI fluency in context.

Develop interfaces or dashboards that allow users to interact with the AI system and adjust its behavior—interfaces can provide checkboxes, controls, sliders, spaces for commonly used prompts or data points, or other mechanisms for users to set preferences, thresholds, or trade-offs. Writer.com's enterprise AI toolset includes interfaces where users can granularly adjust compliance checks, tone, and elements of style—functions that are not easily accessible in tools like ChatGPT Plus (at the time of this publication).

A preferences page in the writer.com app titled 'Clarity,' showing controls for readability level (eg simple, general, advanced) and specific controls like passive voice or wordiness.
Writer.com offers a number of controls to steer the tone and style of their enterprise users' AI instances.

Adobe's Firefly generative AI tools for images create parameter-specific controls that aren't only text-based. While a text-based prompt like 'show a person standing in a room' is a starting point, once the face of a person has been synthesized, specific controls, like how open the eyes are or the kind of smile, are revealed to the user and can be adjusted precisely.

Screen grab of a promotional video from Adobe Firefly product with an AI-generated portrait of a person and a pop-over palette with sliders for various features like smile, age, yaw, eye direction, hair waviness, hair bangs length, eye wrinkles, lighting left, lighting right, and glare reduction.
Adobe's Firefly teaser video showed a Smart Portrait generative AI tool with a hybrid interface. First, it synthesizes a human based on a text prompt, and then allows precise control over various aspects of their appearance using sliders and numbers instead of additional text prompts.

Monitor outcomes

Implement mechanisms to monitor and evaluate the AI system's performance continuously. This includes assessing how well it aligns with the defined goals, constraints, and user feedback.

Monitoring helps identify potential biases, errors, or deviations and allows for timely adjustments.

From an ethical perspective, informed consent, continuous monitoring, and harm prevention/mitigation are a must. This is for moral reasons and because without these functions, user trust will be eroded—especially in so-called 'black boxes' or in moments where unexpected, 'uncanny valley' results occur. In our Data Ethics Guidebook, you can read more about informed consent, doing no harm, data fluency, and harm reduction.

Refine over time

Use the collected feedback, user inputs, and monitoring results to refine and improve the AI system. This iterative process helps align the system's behavior more effectively with the desired objectives.

Steerable AI is imperfect. Here are some tips.

Pay attention to how complex the decision space is

As AI systems become more sophisticated and complex, their decision-making scope expands. Defining all possible rules, constraints, and preferences can be challenging to steer the AI system effectively in complex environments. For example, getting an email-based 'AI assistant' to help you schedule a meeting with just one other person is hard enough (especially if there are conflicting meeting priorities). But getting it to manage entire project timelines for you is orders of magnitude more complex—and has bigger risks, too.

Raise the AI Fluency of your 'steer-ers'

Steerable AI systems often require user input and guidance. However, users may have low or no AI Fluency—the expertise or understanding of the underlying AI algorithms and the implications of their use. Limited AI Fluency will affect your ability to provide accurate guidance or make informed decisions. Cheap or overly-segregated approaches to AI steering—such as using 'click-workers' or 'taskers' who may have limited context or investment in your outcomes—can mean missing important ' big picture' issues. (Read more about building Digital Fluency in general here.)

Align incentives and reconcile trade-offs

Different users or stakeholders may have conflicting goals or preferences. Balancing these trade-offs and resolving conflicts to steer the AI system consistently and coherently can be difficult. Try establishing not just rules but also flexible decision principles—especially 'nesting' decision principles. First, having conversations about objectives that are in tension, even if they are both rated high priorities (such as specificity of answers to users vs. accuracy of answers) will raise important strategy questions. Second, these guidelines for improvisation will prompt critical thinking from humans and may result in better answers from generative AI systems, too.

Make good quality data available to AI systems and teams

Steerable AI relies on data for training, user feedback, and decision-making. Limited, incorrect, out-of-date and/or biased data will affect the system's effectiveness and steerability. If a toddler picks up an adult's iPad and uses the same YouTube account, neither of you will get relevant results (and some may be very inappropriate). All natural language data, like that used by generative AI systems, is biased due to the nature of human brains and human language. Amazon famously terminated a natural language model designed to vet new employee applicants because they could not sufficiently de-bias the system. Even when explicitly stripping gender, race, and class markers from the model, subtle and ingrained language cues were impossible to root out completely. This inherent bias in language is why augmenting human efforts is often easier than automating entire processes to run without some human connection.

Try to use transparent, interpretable systems

The interpretability of AI models is essential for understanding and steering their behavior. However, some AI models, such as deep neural networks and generative pre-trained transformers, can be inherently opaque and lack interpretability. Because it is challenging to comprehend and modify their decisions, you may not be able to engage in 'explainable AI' or 'forensic AI' or otherwise determine the causes of unintended outcomes. Traditional credit scores rely on explicable algorithms, which is why tools like NerdWallet's Credit Score Simulator can tell you quickly what will result in a change to your rating. But novel deep learning or generative AI systems can't always be explained—making them risky to implement in situations where fairness matters deeply. It's one thing to use generative AI to edit marketing copy—it's another thing entirely to ask it to assess a job applicant's cover letter.

Try to foresee unintended consequences, like overfitting

Adjusting an AI system based on user feedback or rules can lead to unintended consequences even while it solves other problems. Changes in one aspect of the system's behavior may affect other areas, and it can be challenging to anticipate and mitigate these ripple effects. For example, quickly adding source data about all of a firm's clients might seem important to tailor the quality of a system, but it may also introduce erroneous data, risk data leakage of personally identifiable information (PII), or cause 'overfitting'—when an AI tool may perform worse when it has too many data points to parse. This is particularly true with generative AIs, which might paradoxically become less 'creative' when over-saturated with data. Like a human with a traditional 'book' education, an overfit system may be prone to recalling from memory rather than generating new results or 'seeing' new opportunities.

Understanding complete data supply chains—including data sources, how data is analyzed and manipulated, and how it is applied—is critical to foreseeing potential negative outcomes.

Make sure you have the resources for steerable AI

Steerable and/or customizable AI systems can use lots of computational (and thus electrical) power, which is expensive and resource-hogging. This can be particularly true when incorporating user feedback or interactive interfaces. At the same time, poorly-steered AI can cause major waste in other ways. Scaling steerable AI systems to handle large-scale applications like company-wide tools or real-time scenarios like live-internet-connected chatbot systems may pose challenges beyond what we can solve today. And users have a funny habit of getting angry when oversold features like steerable AI, only to discover that it is ineffective. Sometimes, reliable access to a generalized AI that is not steerable may be a more realistic expectation to set.

Remember that steerable AI is not automatically ethical and legal

Steerable AI is often associated with ethics, but upvoting or downvoting content isn't limited to harm reduction. Tay, an early chatbot system publicly released on Twitter by Microsoft, devolved into a racist troll in just a few hours because it was being 'steered' by fellow Twitter users. Designers and administrators must ensure that the system's behavior aligns with legal requirements, ethical guidelines, and societal norms.

Striking the right balance between tailoring things to user control and preventing misuse or harm is essential. While not perfect, highly-scrutinized systems with well-funded and well-intended ethics teams, developers, and leaders may not be perfectly adapted to everyone but might be better than open-source or home-grown solutions without comprehensive oversight, especially in the early days of new tech like generative AI.

For example, OpenAI's ChatGPT will try to stop users from accessing information on how to make a dangerous weapon, but a less-governed system might not. The same can be true in reverse—'lowest-common-denominator' answers might not attend to important nuances, such as in responses to questions about rare health conditions.

Keep steerability in mind even if you're not entirely sure how to execute it

All AI strategies should consider AI steerability, explainability, and ethics from day one. While limitations exist, research and advancements in AI are addressing many of the challenges of explainable AI, and AI Fluency is rising. Techniques such as explainable AI, fairness-aware learning, and active learning are being developed to enhance AI systems' steerability, transparency, and accountability. With generative AI, 'grounding' approaches are helping to make sure factual answers are given to human questions of AI systems.

This is one of a multi-part series on Questions to Answer in AI Strategy. MJ Petroni is a Digital Fluency and AI readiness speaker, author, and accelerator. You can read even more in the Digital Fluency Guide.