MIT researchers have developed a new explainable AI method that improves the accuracy and transparency of computer vision models, helping users trust AI predictions in healthcare and autonomous driving.

Researchers at MIT have developed a new approach to make computer vision models more transparent, offering a potential boost to trust and accountability in safety-critical applications such as medical diagnosis and autonomous driving.

In a media statement, the researchers said the method improves on a widely used explainability technique known as concept bottleneck modeling, which enables AI systems to show the human-understandable concepts behind a prediction. The new approach is designed to produce clearer explanations while also improving prediction accuracy.

Why explainable AI matters

In areas such as health care, users often need more than just a model’s output. They want to understand why a system arrived at a particular conclusion before deciding whether to rely on it. Concept bottleneck models attempt to address that need by forcing an AI system to make predictions through a set of intermediate concepts that humans can interpret.

For example, when analysing a medical image for melanoma, a clinician might define concepts such as “clustered brown dots” or “variegated pigmentation.” The model would first identify those concepts and then use them to arrive at its final prediction.

But the researchers said pre-defined concepts can sometimes be too broad, irrelevant or incomplete for a specific task, limiting both the quality of explanations and the model’s performance. To overcome that, the MIT team developed a method that extracts concepts the model has already learned during training and then compels it to use those concepts when making decisions.

The approach relies on two specialised machine-learning models. One extracts the most relevant internal features learned by the target model, while the other translates them into plain-language concepts that humans can understand. This makes it possible to convert a pretrained computer vision model into one capable of explaining its reasoning through interpretable concepts.

“In a sense, we want to be able to read the minds of these computer vision models. A concept bottleneck model is one way for users to tell what the model is thinking and why it made a certain prediction. Because our method uses better concepts, it can lead to higher accuracy and ultimately improve the accountability of black-box AI models,” Antonio De Santis, lead author of the study, said in a media statement.

De Santis is a graduate student at Polytechnic University of Milan and carried out the research while serving as a visiting graduate student at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL). The paper was co-authored by Schrasing Tong, Marco Brambilla of Polytechnic University of Milan, and Lalana Kagal of CSAIL. The research will be presented at the International Conference on Learning Representations.

Concept bottleneck models have gained attention as a way to improve AI explainability by introducing an intermediate reasoning step between an input image and the final output. In one example, a bird-classification model might identify concepts such as “yellow legs” and “blue wings” before predicting a barn swallow.

However, the researchers noted that these concepts are often generated in advance by humans or large language models, which may not always match the needs of the task. Even when a model is given a fixed concept set, it can still rely on hidden information not visible to users, a challenge known as information leakage.

“These models are trained to maximize performance, so the model might secretly use concepts we are unaware of,” De Santis said in a media statement.

The team’s solution was to tap into the knowledge the model had already acquired from large volumes of training data. Using a sparse autoencoder, the method isolates the most relevant learned features and reconstructs them into a small number of concepts. A multimodal large language model then describes each concept in simple language and labels the training images by marking which concepts are present or absent.

The annotated dataset is then used to train a concept bottleneck module, which is inserted into the target model. This forces the model to make predictions using only the extracted concepts.

The researchers said one of the biggest challenges was ensuring that the automatically identified concepts were both accurate and understandable to humans. To reduce the risk of hidden reasoning, the model is limited to just five concepts for each prediction, encouraging it to focus only on the most relevant information and making the explanation easier to follow.

When tested against state-of-the-art concept bottleneck models on tasks including bird species classification and skin lesion identification, the new method delivered the highest accuracy while also producing more precise explanations, according to the researchers. It also generated concepts that were more relevant to the images in the dataset.

Still, the team acknowledged that the broader challenge of balancing accuracy and interpretability remains unresolved.

“We’ve shown that extracting concepts from the original model can outperform other CBMs, but there is still a tradeoff between interpretability and accuracy that needs to be addressed. Black-box models that are not interpretable still outperform ours,” De Santis said in a media statement.

Looking ahead, the researchers plan to explore ways to further reduce information leakage, possibly by adding additional concept bottleneck modules. They also aim to scale up the method by using a larger multimodal language model to annotate a larger training dataset, which could improve performance further.

This latest work adds to growing efforts to make AI systems not only more powerful, but also more understandable in domains where trust can be as important as accuracy.

A major report, Voices for Planetary Health: Leveraging AI, Media and Stakeholder Strengths for Effective Narratives to Advance Planetary Health, produced by the Sunway Centre for Planetary Health at Sunway University and implemented by Internews, offers the first systematic mapping of how planetary health issues are communicated across the world. Its conclusion is clear: ineffective, fragmented communication is undermining humanity’s ability to respond to accelerating environmental and health crises. A Fractured Narrative The research team analysed 96 organizations and individuals across nine countries through interviews and social media mapping. What they found was striking. Despite decades of science showing the deep interconnections between climate change, pollution, biodiversity loss, and human health, global communication remains disjointed, inconsistent, and highly vulnerable to misinformation.

“We know the science. What we lack is a shared story that resonates across communities, cultures, and decision makers,” said Prof. Dr. Jemilah Mahmood, Executive Director of the Sunway Centre for Planetary Health. Most communication efforts are siloed—environment separate from health, climate from social justice, science from lived experience. The report notes that short-term projects, scarce resources, and discipline-bound narratives prevent the creation of powerful, sustained public messages capable of shifting policy or behaviour. AI: Powerful and Dangerous One of the study’s most urgent insights concerns artificial intelligence.

AI can dramatically expand communication capacity through multilingual translation, rapid content generation, and greater accessibility. But it also creates new risks that threaten planetary health messaging. Generative AI tools can be weaponized to fabricate climate falsehoods—from bot-driven denialist content to deepfake campaigns undermining activists. AI systems also reflect structural bias; research cited in the report shows that many models privilege Western epistemologies while marginalizing Indigenous and local knowledge, contributing to what scholars term “global conservation injustices.”

And AI’s own environmental footprint cannot be ignored. Data centres already consume about 1.5 percent of global electricity, with AI-specialized facilities drawing power comparable to aluminium smelters. Training advanced models such as GPT-4 requires three to five times more energy than GPT-3—an escalation that amplifies the very planetary pressures the field is trying to solve.

Communities Most at Risk Are the Least Heard The communication gap most severely harms those already disproportionately burdened by climate-related health threats. The report highlights how marginalized communities—including low-income groups, Indigenous peoples, and communities of colour—face higher exposure to extreme heat, flooding, respiratory illnesses, vector-borne diseases, and pollution-driven health impacts.

These same communities often lack access to reliable planetary health information. Complex scientific jargon, limited translation, and English-dependent messaging create substantial barriers, leaving many without the knowledge needed to advocate for or protect themselves.Multiple studies confirm that racially and socioeconomically marginalized communities in the United States experience greater impacts from climate related health events, including extreme heat, flooding, and respiratory illnesses. Children of colour are particularly vulnerable, experiencing disproportionate health impacts from climate exposures compared to white children. The communication barriers compound these vulnerabilities.

Scientific jargon makes planetary health concepts inaccessible to general audiences, while language delivery challenges—including complex English or lack of translation—further limit reach to non-English speaking communities. Yet young people emerge as a rare bright spot. The study finds that youth activists are using digital platforms— especially Instagram, TikTok, and community networks— to push for environmental accountability. But they still confront algorithmic bias, inconsistent platform moderation, and limited institutional support.

A Blueprint for Coherent, Inclusive Communication

To fix the communication failure, the report proposes a dual framework: strategic communication aimed at policy, and democratic communication rooted in community level dialogue. It outlines six guiding principles: centering marginalized voices; treating planetary health as one integrated story; connecting disciplines and geographies; anticipating backlash and protecting communicators; adapting messages to cultural context; and working with people’s existing mental models. “Communication is not just a tool; it is a catalyst for change.

By speaking with courage, coherence, and compassion, and equipping all actors to tell inclusive stories, we can turn knowledge into action and ensure no voice is left behind,” said Jayalakshmi Shreedhar of Internews. As political rollbacks weaken environmental safeguards and six of nine planetary boundaries are already breached, the stakes could not be higher. Science alone will not save us. A compelling, coherent planetary health narrative—shared across societies—just might

A new study from the University of Oxford has issued a stark warning about the future of global temperatures, finding that nearly half of the world’s population could be living under conditions of extreme heat by 2050. If global warming reaches 2°C above pre-industrial levels—a scenario climate scientists see as increasingly likely—around 3.79 billion people could experience dangerously high temperatures, reshaping daily life across the planet.

The findings, published in Nature Sustainability, suggest that the impacts of rising temperatures will be felt much sooner than expected. In 2010, approximately 23% of the global population lived with extreme heat; this figure is projected to rise to 41% in the coming decades. The study warns that many severe changes will occur even before the world crosses the 1.5°C limit set by the Paris Agreement.

Central African Republic, Nigeria, South Sudan, Laos, and Brazil are expected to see the largest increases in dangerously hot temperatures

According to the study, countries such as the Central African Republic, Nigeria, South Sudan, Laos, and Brazil are expected to see the largest increases in dangerously hot temperatures. Meanwhile, some of the world’s most populous nations—including India, Nigeria, Indonesia, Bangladesh, Pakistan, and the Philippines—will have the highest numbers of people exposed to extreme heat.

The research also shows that colder countries such as the United Kingdom, Canada, Sweden, Finland, Norway, and Ireland could experience relatively dramatic increases in the number of hot days. Compared with the 2006–2016 period, warming to 2°C could lead to a 150% increase in extreme heat days in the UK and Finland, and more than a 200% increase in countries such as Norway and Ireland.

This raises concerns because infrastructure in colder regions is largely designed to retain heat rather than release it. Buildings that maximise insulation and solar gain may become uncomfortable—or even unsafe—during hotter periods, placing additional strain on energy systems and public health services.

Dr Jesus Lizana, lead author of the study and Associate Professor of Engineering Science at the University of Oxford, said the most critical changes will occur sooner than many expect. “Our study shows most of the changes in cooling and heating demand occur before reaching the 1.5°C threshold, which will require significant adaptation measures to be implemented early on,” he said. He added that many homes may need air conditioning within the next five years, even though temperatures will continue to rise if global warming reaches 2°C.

Dr Lizana also emphasised the need to address climate change without increasing emissions. “To achieve the global goal of net-zero carbon emissions by 2050, we must decarbonise the building sector while developing more effective and resilient adaptation strategies,” he noted.

Dr Radhika Khosla, Associate Professor at Oxford’s Smith School of Enterprise and the Environment and leader of the Oxford Martin Future of Cooling Programme, described the findings as a wake-up call. “Overshooting 1.5°C of warming will have an unprecedented impact on everything from education and health to migration and farming,” she said, adding that sustainable development and renewed political commitment to net-zero emissions remain the most established pathway to reversing the trend of ever-hotter days.

Rising temperatures will have far-reaching impacts beyond discomfort. Demand for cooling systems is expected to rise sharply, particularly in regions that already struggle with access to electricity. At the same time, demand for heating may decline in colder countries, leading to uneven shifts in global energy use.

Dr Luke Parsons, a senior scientist at The Nature Conservancy, said the study adds to evidence that heat exposure in vulnerable communities is accelerating faster than previously predicted. He noted that communities least responsible for climate change often face the harshest impacts, underscoring the environmental justice dimensions of the crisis. Addressing the challenge, he said, will require urgent action on both mitigation and adaptation, including rapid emissions reductions and the expansion of equitable cooling solutions.

As global temperatures continue to rise, extreme heat is no longer a distant threat. It is a present and growing challenge that will shape health, livelihoods, and living conditions for billions of people unless decisive action is taken.