August/September 2026
Carbon Vs. Toxicity
Climate policy has a habit of treating greenhouse gases as the only pollution that counts. When a technology can be claimed to cut carbon emissions the industry can use that claim to wave through toxicity, air pollution, and hazardous waste as the price of progress. Examples include recent proposed relaxing of regulations for plastic pyrolysis and chemical-plant air pollution rules. Similarly, the switch from the HFCs-to-HFOs refrigerants swapped strong greenhouse gases for chemicals that degrade into toxic air pollutants.
Plastic Pyrolysis
The American Chemistry Council (ACC), the main lobby organization of the chemical industry, has been an advocate for plastic recycling and pyrolysis, underlining the climate benefits. The ACC backed a recent proposal by the EPA to regulate pyrolysis facilities as manufacturing rather than solid-waste disposal facilities, which would relax the standards for the pollution permitted.
Plastic pyrolysis heats polymers with little or no oxygen, breaking them into oils, gases, char, and mixed organics, some of these then sold as fuels or chemical feedstocks. The industry markets the process as “advanced recycling” and a climate win: keep plastic out of landfills and incinerators, displace virgin fossil feedstocks, count the difference as avoided emissions.
Life cycle assessments show the magnitude of the avoided carbon emissions vary substantially with feedstock, substitution rate, geography, and whether additional treatment is required. Studies also conclude that pyrolysis has a significant impact in several other environmental categories such as acidification, eutrophication, photochemical, and ozone formation. The pyrolysis oil contains harmful polycyclic aromatic hydrocarbons and plants have also reported large hazardous-waste streams.
Chemical Manufacturing
The same ACC recently highlighted ethylene oxide, fluoropolymers, and N-methylpyrrolidone (NMP) as chemicals used in lithium-ion batteries and argued that stricter EPA rules around the production of these chemicals could undercut electrification and the energy transition process. Ethylene oxide is a carcinogen. NMP has documented reproductive and developmental toxicity. Many fluoropolymers are included in the forever chemicals PFAS class.
Along the same lines, in 2025, the EPA exempted chemical plants producing synthetic organic chemicals and polymers from a 2024 rule that regulates air emissions from chemical plants.
HFOs as a Regrettable Substitution.
High global warming potential Hydrofluorocarbon gasses (HFCs) are being replaced with Hydrofluoroolefins (HFOs) that have much lower global warming potential.
HFO-1234yf widely used in car air-conditioning degrades in the atmosphere to trifluoroacetic acid (TFA). The European Chemicals Agency classified TFA as a reproductive toxicant, as well as persistent and mobile, earlier this year.
Hydrocarbons, carbon dioxide, ammonia, and other HFOs with lower TFA yields can be used in some equipment. Unfortunately, many mobile-air-conditioning systems were designed around HFO-1234yf.
Hazard and pollution must be considered alongside climate before chemical substitutes are commercialized.
The Perfect Data Strawman
There’s a nagging disconnect between how we behave as individuals vs. how we behave collectively regarding toxic chemicals.
As individuals we are willing to go to great lengths and expense to avoid the slightest exposures to suspected toxicants. The decision algorithm is instinctual and simple, everything suspected of being potentially poisonous should not be touched, just like any mushroom or berry in the forest we don’t know for sure is safe. We often practice an innate precautionary principle.
Collective decision-making is more complicated: it tries to weigh risks against benefits, but often moves slowly. Regulators assess both a chemical’s hazard (the adverse effects it can cause) and potential exposure (how much people encounter, how often, by what route, and which populations are affected) to produce a risk characterization. This process is laborious and often favors incumbent chemicals, because approving safer alternatives or removing existing substances can take years.
For most chemicals, we will never have precise dose-response curves covering every route of exposure, every stage of life, every mixture, and every person. It is important to understand the limits of toxicology, so we do not fall into the trap of asking for perfect knowledge before we act. Here are some of the reasons why.
Limitations of animal studies - toxicology relies on animal studies, but translating results between species is imperfect. Chocolate is harmless to humans and lethal to dogs while thalidomide or celecoxib passed animal testing before issues were seen in humans. Different species absorb, metabolize, distribute, and eliminate chemicals differently. Receptors and biological pathways can also vary. Scaling factors and safety margins attempt to bridge the gap, yet they remain models with limitations.
Acute toxicity and chronic toxicity are different questions - the classic LD50 acute toxicity test measures the dose that kills half a test population. That says little about chronic, low-level exposure. Sugar is one of the safest substances from an acute standpoint, but decades of regular intake can cause metabolic disease. Certain chemicals, such as endocrine disruptors, have low-dose activity or activity that is not linearly linked to the dose (“non-monotonic”). High-dose testing alone cannot answer every question about lower, real-world exposures.
Real life exposure is never in isolation - while most toxicology studies test one substance at a time, people absorb hundreds of compounds daily, some of which may interact additively, antagonistically, or synergistically. Testing every plausible combination is impossible, leaving mixture effects poorly mapped.
Age - developing bodies are often more sensitive than adults because their organs are still forming, they have different metabolic pathways, different blood-brain barriers, different rates of cell division, less mature detoxification pathways. EPA specifically recognizes that exposure during critical windows of development can produce effects that would not occur at the same dose later in life. FDA recognizes the same thing as the drug approval process is different for children and teens compared to adults.
Biological sex - reproductive and hormonal systems differ substantially between the biological sexes, and many endocrine-disrupting chemicals are specifically hypothesized to act on those systems.
Individual variation - differences in metabolizing enzymes, receptor shapes, nutritional status, prior exposures, gut microbiome, pre-existing conditions, and other factors mean two people with identical exposure histories can have different outcomes. Some lifelong smokers never develop lung cancer while others develop it from little exposure. This variability is why population-level statistics do not predict an individual outcome well.
Mechanism of action - even when associations between exposure and health outcomes are visible in epidemiological studies, the underlying mechanism of action — how exactly a chemical disrupts a cellular process — is often not established because that requires expensive, multi-year research. The absence of a fully mapped molecular mechanism does not by itself demonstrate that an epidemiological association is false.
There is no definitive human experiment - an experiment that would expose one large, randomized human population to a realistic low-dose chemical combination for long periods of time, and compare against an unexposed control group, and control for every confounder — cannot be designed. It would be unethical to deliberately expose people, and practically impossible to isolate individuals from all outside contamination.
Other factors - include long latency periods (effects may appear decades after exposure), new chemicals are entering commerce faster than comprehensive testing takes place, and analytical chemistry having limits (measuring trace concentrations in biological systems is expensive). Safety testing is often paid for by the industries with an interest in the outcome so funding and regulatory incentives shape which questions get asked and how thoroughly they are pursued. Independent, well-funded, mechanism-level research is rare and slow.
Perfect data is not coming - it would be unreasonable to ban a chemical the moment a concerning study appears, as restrictions on important chemistries can carry economic and social costs. But it is just as unreasonable to demand definitive proof before acting. Waiting for the perfect data can become a permanent excuse for inaction.
How much evidence should we require before changing course? That answer should depend on the seriousness of the suspected harm, strength and consistency of the evidence, degree of exposure, persistence and bioaccumulation, availability of safer alternatives, and consequences of being wrong in either direction. It is indeed difficult and complicated to navigate these kinds of decisions.
Where safer alternatives are available the course of action is simpler. If different chemicals perform a similar function at comparable cost, but one has substantially less evidence of persistence, bioaccumulation, endocrine activity, carcinogenicity, or developmental toxicity, we should not need full human studies to overcome inertia and adopt it.
The goal is to design chemicals and materials with lower inherent hazards, improve toxicity testing, avoid regrettable substitutions, and replace problematic chemistry when the weight of evidence and the availability of alternatives justify it.
Can We Turn Compute into Safer Chemistry?
Attention and funding for artificial intelligence in chemistry has focused on drug discovery, with billions invested to predict protein structures, identify therapeutic targets, and design new molecules. There are three ways new computational tools can be used to develop safer chemistry: 1. to organize the knowledge we already have about chemical hazards; 2. to predict potential toxicity; and 3. to make experimentation and formulation more efficient.
1. Chemical hazard information is fragmented. Information about chemical hazards exists across multiple sources, such as government agencies’ data, scientific papers, safety data sheets, toxicology studies, and restricted-substance lists. Finding, curating, and interpreting this information has traditionally been expensive.
AI agents have been taught to identify chemical synonyms, search the literature, pull regulatory classifications, extract study results, compare evidence across endpoints, and assemble a preliminary hazard assessment for a certain chemical. The most promising tools combine language models with curated databases, cheminformatics tools, predictive models, and traceable source material. Companies that have developed agents for chemical hazard assessment include Human Chemical and Clearya. Human Chemical provides an output similar to a toxicologist, while Clearya searches for potential chemicals of concern in large sets of consumers, regulatory, or financial data. Traditional sources of curated chemical hazard information have a significant role to play. Data bases like ChemForward, Pharos, and SciveraLENS, complement the AI platforms.
2. Predicting Toxicity. There are two approaches to predicting toxicity, the first relies on modeling small molecule - protein interactions, the second relies on correlating chemical structure with toxicity.
AlphaFold 3 is a Google developed platform for modeling small molecule - protein interactions and can be used to predict toxicity. Isomorphic Labs and SandboxAQ are private platforms doing the same thing but focused on drug discovery. They could also be used for predicting toxicity based on modeling protein interactions.
ProTox 3.0 and FDA’s SafetAI both predict toxicity based on chemical structure but rely on large datasets of chemical toxicity that are not currently available for all types of toxicity and classes of chemical structures. The FDA initiative aims to develop more complete data and new toxicity models for hepatotoxicity, carcinogenicity, mutagenicity, nephrotoxicity, and cardiotoxicity.
Neither approach is sufficient alone. Interaction models can miss chemistry that never binds a well-modeled protein. Structure–activity models can fail on novel chemical structures. Used together, they give chemists a way to reject potential toxic chemicals and focus on the ones that are expected to be safer.
3. Run fewer, better experiments. Even good predictions leave a gap that only measurement can close and machine learning can help decide which experiment is worth running next. Uncountable, Citrine Informatics, Turing Labs, and Albert Invent all combine proprietary data sets, molecular-property prediction, and design of experiments algorithms to speed up the discovery of new materials and formulations. For safer chemistry these tools can enable multi-objective design: performance, cost, processability, and lower hazard in the same search. That only works if hazard is an explicit target in the optimization, not a compliance review after the fact.
The key constraint is data, not models. The limitation across all computational tools is the access to large, high-quality datasets. Many industrial laboratories have data from decades of experimental results, but this data is stored in PDFs, spreadsheets, handwritten notebooks, paper, and the experience of individual chemists. Failed experiments are particularly important and are less likely to be included in datasets. Most chemistry AI companies rely on the same sources including PubChem, ChEMBL and DSSTox/CompTox + ToxCast (toxicity). These are okay starting places but still lack experimental details and toxicity attributes and often the volume and quality of data varies wildly for different chemicals.
Digitizing chemistry data and developing new tools to cheaply and effectively generate new data is likely more important for chemistry AI than developing new models.
Financings
Aardaia breeds nitrogen-fixing potatoes, and raised $5.7 million. Apoha generates data on how molecules and materials behave in liquid samples, with applications in drug discovery, and raised $36 million. Birdsview improves the inspection of concrete structures for safety, and raised $4.3 million. Certo uses AI agents to verify product compliance across ingredients, formulas, labeling, claims, and market-specific regulations, and raised $4 million. Circular Materials recovers critical raw materials from industrial wastewater, and raised $13.5 million. Claros Technologies destroys PFAS chemicals in industrial and contaminated ground water, for manufacturers, utilities, remediation firms, and government agencies, and raised $55 million. CuspAI helps discover materials that reduce research times and reliance on rare metals, and raised $450 million. Eurodia Industrie builds industrial liquid-treatment systems that recover valuable materials and reduce waste from liquid streams, and raised $20.5 million. Fast Metals extracts minerals from aluminum-refining red mud using chemicals from refinery waste, and raised $4.3 million. GR3N breaks down PET plastics and polyester waste using a microwave process and raised $17.9 million. Kind Designs builds 3D-printed seawalls and shoreline structures that protect coastlines while creating habitat for marine life, and raised $10 million. Moa Technology develops synthetic and biological herbicides, and raised $29.6 million. Lium structures complex scientific and industrial datasets into formats large language models can query to generate consistent analytical outputs and raised $5.5 million. Mirendil builds AI tools that help scientists create specialized models for medical and materials research and raised $200 million. New Dawn Bio grows shaped wood products from tree stem cells to reduce material waste and production costs and raised $2.4 million. Quercus Biosolutions designs peptide crop protection compounds that control herbicide-resistant weeds and other pests and raised $5 million. Red Metals develops integrated copper refining and manufacturing operations that convert scrap feedstocks into finished copper products, and raised $10 million. Sonic Fire Tech makes acoustic fire suppression systems that use infrasound to extinguish fires without water or chemicals, and raised $15 million. |
