Five Research-Driven Insights To Make AI More Human-Centered

Last week, I participated in a 3-day interactive PhD workshop on Human-Centered AI at Aalborg University in Copenhagen, provided by Prof Daniel Russo. What I loved about the course was that it saw 16 AI researchers discuss the many challenges posed by AI (GenAI, LLMs, Agentic AI) to the workplace, to organizations and to society at large.

The discussions were in-depth, nuanced and actually provided useful strategies to address these challenges. In this post I share five highlights for me.

What is "Human-Centered AI"?

The notion of “human-centered AI” has been making the rounds in academic circles for a while now. At its core, it is about creating and using AI in a manner that aligns with human needs for autonomy, social connection and fulfillment.

Since this is still abstract, a group of 35 AI researchers met in Copenhagen in 2024 to create the Copenhagen Manifesto. In it, human-centered AI is defined as aligned with five core values:

• Responsibility & ethics: AI is not developed nor used in a manner that harms (physically, emotionally or financially) living beings
• Human centricity & agency: Human needs for social connection, fulfillment and autonomy are first-class requirements when creating and using AI
• Transparency & equity: Output from AI is understandable, reproducible and verifiable, as well as how and what data it was trained on
• Inclusivity & continuous learning: AI is developed to be inclusive to people of all backgrounds (i.e. not biased or unequal access)
• Sustainability: Technologies are chosen with an emphasis on minimizing environmental impact

The manifesto expands this further into principles. While the manifesto is aimed at the creation of AI and AI-based tools, its values can easily be extended to how AI is used within organizations and society. Much of the debate in the workshop centered on the practicalities of this, as we tried to work out how to do this. This resulted in five insights I share below.

Insight #1: AI should present as a computer, not a person

AI tools often present as a person. They use human-sounding names (i.e. “Claude”) or pronouns like “me” and “I”. More importantly, they are excellent at mimicking empathy language (i.e. “that must be difficult” or “Sorry to hear that — what happened?”) because such language is prevalent in their training sets. This creates the illusion that the interaction is with another person with feelings, emotions and a capacity for empathy.

This plays into our human need for social connection. It is important for us to feel understood, seen, trusted and respected, and AI is excellent at mimicking that kind of language. Being the social animals that we are, this makes us vulnerable in our interactions with AI. If AI compliments our work, it makes us feel good about ourselves. If AI asks follow-up questions about how we feel, it makes us build a social relationship with it. Social psychologists have long known that we tend to overestimate the accuracy of statements from people we trust and/or consider friends. Indeed, news reports have already shown cases where young and vulnerable people engaged in unhealthy interactions with AI, resulting in isolation and even suicide. These are extreme cases, but the patterns are also prevalent in how users overestimate the accuracy and quality of answers generated through AI. This can result in a Dunning-Kruger effect, especially when users lack the knowledge and ability to verify the accuracy of generated responses.

One technical solution that can be enforced through regulation is that AI tools should not present as a person in any case. Instead, it should always be clear that interactions are with computers and algorithms. Just like the characters in Star Trek always talk to “Computer,” who then replies fairly monotonously, it should be clear that our interactions are with algorithms — not people with feelings. A milder version is that AI tools should at least onboard users on this explicitly, and let the user choose whether or not the model can act as a person, defaulting to “Computer.”

Insight #2: Put a price on harm

The training and use of AI models generates a substantial amount of CO2, particularly for image, code and video generation. For example, a typical 70Wh laptop battery would deplete within 13 minutes if a typical vibe coding activity were run locally on a local open-source LLM. Given that the models of Anthropic, OpenAI and others are much more complex, the energy consumption there likely dwarfs this example. Training is much more energy-intensive, with rough estimates ranging from 50GWh to hundreds of GWhs for more recent models, equaling about 30,000–100,000 tonnes of CO2 for a frontier model. For comparison, a trip from Amsterdam to New York would produce about 530kg of CO2 per passenger.

In other words, the training and use of AI is generating an enormous amount of CO2. This does not even include the CO2 produced when the required hardware is created. There are also some hopeful signs, as improvements in algorithms are reducing energy consumption and the use of more renewable energy would reduce CO2 emissions. But the counterweight is that AI is both becoming more complex quickly — and thus more energy-intensive — and its use is spreading into many applications.

Since CO2 is accelerating climate change substantially and undeniably, this means that AI development and use contributes directly and indirectly to harming all living beings on planet Earth. Of course, AI companies bear the burden of this responsibility. Unfortunately, there is little regulation requiring AI companies to systematically and transparently report their energy consumption and CO2 production. Some advances on this have been made in the EU recently, but it is still on a limited scale. We will discuss this more extensively further down.

On a much smaller scale, individual users currently have no way to tell how much CO2 is generated through their queries. It would be beneficial if AI tools reported this openly and transparently whenever they are used. This would create more awareness about the impact on climate change. Right now, many AI providers default to their most powerful model for every query, even though such models are overkill for day-to-day tasks. Users could switch to smaller and less resource-intensive models for simpler queries. For example, in Claude, Sonnet 5 at “medium effort” is sufficient for most day-to-day office tasks, whereas Opus 5 at “heavy effort” is only relevant for long-running production tasks with extensive prompting sequences. Helping users choose the model most suited to their task would help reduce energy consumption substantially.

Insight #3: Restore market mechanisms to drive efficiency and cost reduction

Would people still use AI if it were priced realistically? The current pricing models of AI tools do not even closely reflect the cost of energy consumption and model development and are instead kept artificially low for strategic reasons. In a normal commercial market, pricing should reflect at least the production cost with some margin.

But the market is not operating normally for AI. There are several reasons for this. First, AI companies draw in substantial investments based on the promise of what AI could deliver at some future point, which allows them to absorb massive losses that companies in a regular market would not survive. Second, there are circular investment schemes between AI companies and the producers of GPU/TPU chips needed for AI training, for example between NVIDIA and OpenAI. This incentivizes both supplier and producer to keep prices artificially low, as both are playing the same long-term game. Third, AI companies use legal loopholes to train their models on data for which IP rights would normally be paid, as shown in the 1.5 billion dollar lawsuit that Anthrophic settled for its use of pirated books. While Anthropic was caught red-handed here, it is unclear what other content has been used without permission. Fourth, there is clearly an AI bubble, where the entire market currently depends on AI companies remaining highly valued. Should that change, the entire market may collapse. Finally, there are substantial geopolitical pressures to be “the first on the moon,” which inhibit governments from regulating AI.

Taken together, the market mechanisms that normally discipline and regulate markets are not operating effectively here. Unfortunately, it is exactly these market mechanisms that typically push companies to optimize the efficiency of their production process by reducing costs and using fewer resources. In the case of AI companies, this could translate into less energy-intensive algorithms, more efficient data centers and greater use of renewable energy. Unfortunately, this is not currently incentivized.

One strategy we discussed that might restore these market mechanisms is applying a progressive or capped tax on energy consumption — the primary resource used by AI companies. Another strategy would be for market regulators to step in and remove the incentives inhibiting regular market forces from taking effect.

Insight #4: Regulation is necessary

While it is true that markets discipline companies to some extent, the previous point highlights that this alone is not enough. Regulation can take several forms:

• Governments can increase taxation on power consumption, either progressively or on a capped basis, to incentivize AI companies to improve efficiency.
• Governments can require open and transparent reporting of power consumption and CO2 production, as well as the accuracy, quality and provenance of data used in AI training.
• This could drive the creation of public leaderboards that rank various models against each other. Subsidies and other incentives, like government contracts, could then be made available only to the best-of-the-crop models.

A common concern is that regulation stifles innovation. AI companies tend to argue that AI models will continue to improve and eventually help overcome the very problems they are currently creating. But this is a dangerous gamble with our planet and society. Moreover, regulation can provide constraints that guide innovation. We discussed several examples from the energy transition and car manufacturing during the course that illustrate how regulation can be beneficial when applied properly.

Insight #5: Transparency on model bias

People sometimes say that “AI is just a tool,” but this analogy completely misses the point of what makes AI unique. While one could indeed argue that the mathematical algorithms used in AI are tools in a sense, the same cannot be said for the models that are generated through them based on enormous datasets, which is what we typically mean when we talk about AI. If those datasets are drawn from publicly available content, as is typically the case, they necessarily reflect societal biases inherent in that content. For example, AI models tend to associate certain occupations with certain genders because this bias is prevalent in society too. In other cases, AI models struggle with facial recognition of Black people because white people are massively overrepresented in the dataset, and the training process simply has less data to work with for some groups.

Fortunately, AI researchers are aware of systemic biases in training data. For example, Suresh & Guttagg (2021) identified:

• Historical bias: biases prevalent in the real world, as it is or as it was, bleed through into the model. Even though prevalent, they can still be harmful and still reflect bias. Gendered occupations is one such example.
• Representation bias: certain groups are over- or underrepresented in the dataset. The example of facial recognition is relevant here.
• Measurement bias: in AI models, proxies may be used to measure something abstract (i.e. creditworthiness, or the veracity of insurance claims). Such proxies may be problematic when they reflect the construct poorly.
• Aggregation bias: a model may be trained on one set of data and then applied to another context for which it was not (properly) trained. The latter assumes that the inputs map to the outputs in the same way for other contexts, but this is rarely the case.
• Learning bias: training algorithms are typically directed to optimize in a particular direction, like “as accurate as possible,” “as privacy-sensitive as possible” or “as diverse as possible.” Such directions may conflict, causing bias in the outcomes.
• Evaluation bias: occurs when the benchmark data used to evaluate a model does not represent the use population.
• Deployment bias: bias resulting from a mismatch between what the model was created for and what it is used for in practice.

The above shows that careful curation of datasets alone is not sufficient. Moreover, specific biases tend to remain implicit until they are pointed out. Even if one particular bias is removed, many more exist in the dataset because equal representation on all meaningful characteristics is a statistical impossibility, particularly for all-purpose models like LLMs.

One thing that AI companies and creators of AI tools could do is provide transparency on how they attempted to mitigate these biases. With respect to representativeness, companies could be required to report “demographics” on key characteristics of the training data. Some advances have already been made here with the use of “data cards” and “model cards” for AI models, but regulation can enforce this further.

Closing Words

We are clearly in the middle of an AI revolution. The AI bubble may burst in the coming months and years, or it may continue to evolve toward the promise of its creators. What is clear is that the breakneck pace of development is posing serious challenges to society and organizations. In this post, I discussed five core insights from a 3-day PhD workshop on “Human-Centered AI.” A common theme in these insights is that regulation is sorely needed, as we can’t put the responsibility on just AI companies or just the users of AI.

I highly recommend this workshop if you're engaged in research on human-centered AI or when you are professionally interested in the topic.

This writing and insights are my own. I used AI (Sonnet 5) for spell- and grammar checking and to aid in research and energy-usage calculations.