Inside AIXPERT’s research on bias and fairness in multimodal AI
Can an AI system be highly accurate yet fundamentally untrustworthy? New research from AIXPERT suggests the answer may be yes. From text-to-image generation to multilingual reasoning and multimodal foundation models, researchers are developing innovative benchmarks that reveal how today’s AI systems perform not only technically, but also socially and ethically.
As multimodal artificial intelligence systems become increasingly embedded in everyday decision-making, researchers are beginning to confront a difficult reality: these models are not only learning how humans communicate and interpret the world. They are also inheriting many of society’s biases along the way.
From recruitment platforms and healthcare assistants to industrial automation systems and AI-generated media, multimodal models are now capable of jointly processing text, images, audio, and video with remarkable sophistication. Yet beneath this technological progress lies a growing concern within the AI research community: how do we ensure that these systems remain fair, explainable, and trustworthy when they reason about people, professions, and cultures?
This question sits at the centre of several recent scientific studies produced within the AIXPERT Project. Through a series of benchmarks and evaluation frameworks, AIXPERT researchers are examining how multimodal foundation models behave when exposed to demographic and social cues such as gender, race, occupation, age, and language.
Bias in AI-generated images and occupational representation
One line of research focused on text-to-image generation systems and how they portray socially significant professions including CEOs, nurses, software engineers, teachers, and athletes. In Prompting Away Stereotypes? Evaluating Bias in Text-to-Image Models for Occupations (Raza et al., 2025), researchers benchmarked both commercial and open-source models such as DALL·E 3, Gemini Imagen, Stable Diffusion XL Turbo, FLUX, and Grok-2 using hundreds of occupation-related images generated under different prompting conditions.
What emerged was a striking picture of how deeply societal stereotypes can become embedded inside generative AI systems. Certain models consistently associated leadership and technical professions with male or White representations, while care-oriented professions were frequently feminised. Even more interestingly, the biases were not uniform across systems. Different model architectures, from diffusion-based models to autoregressive and hybrid approaches, displayed distinct representational patterns, suggesting that bias propagation is closely tied to how multimodal systems are trained and optimised.
The researchers then explored whether prompt engineering could mitigate some of these imbalances. By introducing fairness-aware prompts explicitly encouraging demographic diversity, several models significantly shifted their outputs. In some cases, this produced more balanced portrayals. In others, however, the systems overcorrected dramatically, replacing one demographic skew with another. As Raza et al. (2025) argue, fairness cannot be treated as a superficial prompt-level adjustment. While prompting can influence outputs, the effects remain unstable and highly model-dependent, requiring deeper interventions at the model and dataset level.
Evaluating fairness beyond accuracy
At the same time, another AIXPERT-related study turned attention toward vision-language models and the role visual context plays in shaping AI reasoning. In Bias in the Picture: Benchmarking VLMs with Social-Cue News Images and LLM-as-Judge Assessment (Narayanan, Khazaie & Raza, 2025), researchers developed a benchmark containing more than 1,300 real-world news images paired with open-ended questions to investigate how multimodal systems interpret social cues such as age, race, gender, occupation, and sports affiliation.
The results exposed a subtle but critical tension within modern multimodal AI systems. Models that performed strongly in terms of factual accuracy and visual grounding were not necessarily the least biased. In several cases, systems produced technically correct answers while simultaneously injecting stereotypical assumptions into their reasoning processes, particularly around gender and occupation. This finding challenges a common assumption in AI development: that increasing model capability automatically improves trustworthiness.
The study also introduced an important methodological contribution through the use of an “LLM-as-judge” evaluation framework. Instead of relying solely on traditional accuracy metrics, Narayanan, Khazaie & Raza (2025) jointly assessed answer relevance, faithfulness, and demographic bias using structured evaluation rubrics. This reflects a broader shift in explainable AI research toward more holistic trustworthiness evaluation methodologies capable of capturing both technical performance and social implications.
The multilingual fairness challenge
The challenge becomes even more complex in multilingual and multicultural settings. In LinguaMark: Do Multimodal Models Speak Fairly? (Raza et al., 2025), researchers evaluated multilingual multimodal reasoning across eleven languages using image-text tasks measuring answer relevance, faithfulness, and demographic bias. The benchmark revealed that many multimodal systems still perform substantially better in English than in lower-resource languages, while fairness inconsistencies become more visible across diverse linguistic contexts.
The LinguaMark study also demonstrated that demographic bias is not distributed evenly across attributes. Gender-related bias consistently remained among the strongest recurring issues across evaluated models, while performance variations across languages highlighted how multimodal AI systems can struggle to generalise fairly across cultural contexts. These findings are particularly relevant for Europe’s multilingual digital ecosystem, where trustworthy AI systems must remain robust and equitable across different linguistic communities.
Toward unified and multimodal evaluation
Yet measuring individual bias dimensions in isolation leaves a deeper question unanswered: how do fairness, ethics, empathy, and robustness interact when evaluated together across realistic conditions? In Humanibench: A human-centric framework for large multimodal models evaluation (Raza, Narayanan, Khazaie, Radwan et al., 2025), researchers constructed a benchmark of approximately 32,000 expert-verified image-question pairs spanning seven human-centred principles across real-world news imagery, and evaluated 15 state-of-the-art models under the same unified framework. What emerged challenged a widespread assumption in the field: that stronger models are more trustworthy models. Although many systems achieved high conventional accuracy, they consistently underperformed on ethical, inclusive, and empathic dimensions, with race-related cues remaining the most persistently difficult attribute across the board.
A companion study extended this scrutiny into native audio-video understanding, a modality that most existing benchmarks had treated as secondary or optional. In SONIC-O1: A Real-World Benchmark for Evaluating Multimodal Large Language Models on Audio-Video Understanding (Radwan, Raza et al., 2026), researchers evaluated models across approximately 60 hours of real-world conversational video spanning 13 domains, from medical consultations and courtroom proceedings to mental health counselling, finding demographic disparities of up to 21.4% on temporal localization even when multiple-choice accuracy appeared stable across groups. Together, both studies point to the same underlying problem: aggregate performance figures actively conceal where models are failing people, and evaluations that do not disaggregate across principles, modalities, and demographic groups will always produce a false sense of progress.
Building human-centred and trustworthy multimodal AI
Collectively, these studies contribute directly to AIXPERT’s broader objective of developing explainable, accountable, and human-centred AI systems. Rather than treating trustworthiness as a downstream moderation layer, the project is investigating how explainability, transparency, accountability, and human oversight can be embedded directly into multimodal and agentic AI architectures themselves.
This includes the development of multimodal trustworthiness benchmarks, explainability assistants, fairness-aware evaluation frameworks, governance mechanisms, and human-in-the-loop methodologies capable of monitoring how AI systems reason and make decisions under real-world conditions. Importantly, these concepts are not being explored solely in laboratory settings. AIXPERT validates its technologies across practical pilots in healthcare, recruitment, manufacturing, educational robotics, and the creative industries. These are domains where hidden bias or opaque reasoning can have direct societal, ethical, and operational consequences.
As multimodal AI systems continue to evolve toward increasingly autonomous and agentic behaviour, the challenge facing researchers is no longer simply building more capable models. Increasingly, the central question is whether these systems can reason in ways that remain transparent, socially aligned, and worthy of public trust.
The research emerging from AIXPERT suggests that achieving trustworthy multimodal AI will require far more than larger datasets and stronger computational performance alone. It will demand new evaluation methodologies, deeper explainability frameworks, and a more nuanced understanding of how multimodal AI systems interpret people, language, and society itself.