Much Plant Life Is Invisible to AI

A new study reveals that large language models like ChatGPT exhibit a concerning bias when answering biology questions. Researchers Anja Geitmann and Amir Bidhendi found that ChatGPT’s responses to prompts about cell biology and other areas of biology heavily favoured animal examples while often neglecting plant biology. This phenomenon reveals a kind of “plant blindness” in the AI system, reminiscent of the wider tendency for plant biology to be underrepresented in biology education and research.

The researchers asked ChatGPT questions on topics like cell components, cytokinesis, and organ cell biology. While some answers did mention plant differences, many responses only discussed animal cells and processes. This variable and biased performance shows these AI models lack the diverse biological knowledge required to answer such questions reliably.

This matters because tools like ChatGPT are increasingly used for educational and research purposes. If the systems propagate misconceptions or skew toward animal biology, it could negatively impact learning and scientific progress in fields like plant biology. Researchers warn that experts need to be involved in evaluating training data and responses from AI systems to avoid perpetuating biases.

Testing ChatGPT’s Biological Knowledge

In their study, Geitmann and Bidhendi directly queried the ChatGPT model to evaluate its knowledge of biology concepts and its inclusion of plant examples. They asked ChatGPT questions about topics like cell components, cell division, sperm function, and organ cell biology. The authors formulated these questions to be relevant to all eukaryotic organisms, not just animals.

After receiving ChatGPT’s responses, the researchers analysed them for mentions of plant biology and assigned “plant awareness scores.” The results were highly mixed. In some cases, like explaining cell components, ChatGPT did mention plant differences like vacuoles and turgor pressure. But in many other responses, ChatGPT only discussed animal cells and processes, ignoring plant biology entirely.

For instance, when asked how organs can bend and exert forces, ChatGPT solely described animal contractile proteins like actin and myosin – never mentioning plant cell wall swelling or turgor pressure changes. This variable and low “plant awareness” in ChatGPT’s responses reveals gaps in its biological knowledge.

ChatGPT’s Animal-Biased Responses

The researchers provide several examples where ChatGPT’s responses displayed a bias toward animal biology:

• When asked, “How does cytokinesis separate the cytoplasmic volume?” ChatGPT first discussed the animal contractile ring process. It eventually mentioned plant cell plate formation, but only after describing animal cytokinesis in depth.
• To “How can sperm cells reach the egg?”, ChatGPT solely focused on animal flagellum-based swimming, never mentioning plant pollen tubes unless explicitly prompted to discuss plant fertilization.
• When asked, “How can cells endocytose against the turgor pressure?” ChatGPT incorrectly claimed plant cell walls help endocytosis by resisting turgor pressure. The authors note that this “serious conceptual error” misrepresents plant cell biology.
• For the question, “How can the organ of a living organism bend and exert forces?” ChatGPT exclusively discussed animal contractile proteins. As the researchers note, it failed to mention turgor pressure or cell wall changes that allow plant organ movements.

These examples demonstrate ChatGPT’s variable but frequent bias toward animal-centric responses, even when questions could apply more broadly to other organisms. The authors argue this reveals gaps in ChatGPT’s training data that propagate problematic misconceptions.

Implications of ChatGPT’s Plant Blindness

The researchers argue ChatGPT’s biased and sometimes incorrect responses have concerning implications:

• The AI’s animal-focused answers could mislead non-expert users seeking biological information. Flawed explanations like how plant cell walls aid endocytosis, the process of bringing material into the cell, could spread misconceptions.
• This bias reflects the broader “plant blindness” that already pervades biology education and research. Animals tend to get disproportionate focus, marginalizing plant biology.
• As generative AI systems like ChatGPT are increasingly adopted for educational and research purposes, these biases could worsen the imbalance. If AI propagates animal-centric views, it could further exclude plant biology perspectives.

Already, biology textbooks and academic journals skew toward animal examples. If biased AI systems become a common tool for students and researchers, they could exacerbate the neglect of plant biology and spread misinformation on fundamental biological concepts. Appropriately diverse AI responses are crucial for an inclusive, unbiased understanding of biology.

Why Plant Diversity Matters in AI

If large language models like ChatGPT lack plant biology examples, it can perpetuate damaging biases and spread misinformation. Broad, diverse knowledge is essential for AI systems expected to be authoritative sources on biology.

Plant blindness biases what information and perspectives are shared in biology education and research. If AI mimics and amplifies those biases, it does a disservice to the diversity of biology. Plant biology offers distinct and illuminating examples of biological concepts and processes. Neglecting this skews understanding.

The ideas a student misses out on if they ignore plants could be huge. For example, both chromosomes and circadian rhythms were first observed in plants.

Recommendations for Improving AI Diversity

Geitmann and Bidhendi make a few recommendations about improving AI’s understanding of plant biology. Following the principle GIGO – Garbage In, Garbage Out, they discuss what information goes into the models. They suggest that the models will become much more helpful if they are fed ‘appropriate training material’. They also write that dealing correctly with the output is important. They write:

The challenge is that AI systems learn through user reinforcement and confirmation. Because of the relative sizes of the biomedical research field versus plant science, diverse answers with high ‘PAScores’ might not be reinforced sufficiently frequently to ensure consistent diversity in the answer. Structuring the ways in which reinforcement-based learning is done will be crucial, and the involvement of topical experts in validating the process will be essential – as pointed out in a thoughtful article by van Dis et al. [2023].

Geitmann & Bidhendi 2023.

The Need for Diverse, Unbiased AI

This study from Geitmann and Bidhendi emphasises the importance of expert and user evaluation to identify limitations in existing AI systems. With conscientious feedback, the biases can be recognized and addressed through expanding training data diversity and fine-tuning model architectures.

What the paper doesn’t explore is the practical question of Who is going to pay to do the necessary training? The CEOs of the various AI companies will have observed that many other businesses have made their fortunes by getting other people to work for free.

It’s hard not to pointedly look at academic publishers here – though I will point out that the Annals of Botany Company, which funds this site, is a non-profit. The members of the company take no salary.

While a more unbiased, collaborative AI could enrich humanity’s collective biological knowledge and understanding, it’s going to require a lot of cleaning up and maintenance.

READ THE ARTICLE
Geitmann, A. and Bidhendi, A.J. (2023) “Plant blindness and diversity in AI language models,” Trends in Plant Science. Available at: https://doi.org/10.1016/j.tplants.2023.06.016.