1. What problem does this essay raise?

The essay is not asking how useful AI is. It asks by what standard anyone should decide whether AI may advance at all. The author lists four concerns that have grown with the spread of generative and agentic AI: that the technology could replace human contribution, override human supervision, ignore ethical and technical guardrails, and cause increasingly irreparable harm to the environment[1]. The author treats these as concerns that could lead to destructive outcomes for life on Earth.

The author, Yvette Running Horse Collin, is executive director and principal science officer of Taku Škan Škan Wasakliyapi: Global Institute for Traditional Sciences (GIFTS)[2]. The piece was published in Science as a one-page essay. It is not a research article reporting experiments or survey data, and the bibliographic record does not say whether it went through external peer review. What follows reads the author's argument and the steps that lead to it.

The question matters to readers in pharma because drug development and regulation already ask "may we use this?" before "can we use this?". The essay shows what happens when that question is posed from a premise other than weighing benefit against harm.

2. What does it argue?

The core claim fits in one sentence: if minimizing harm to life is a condition of science, AI as presently configured does not meet it[1]. In a Lakota context, the author writes, science must ensure as little harm as possible is done to life. From that premise, the mere possibility of the outcomes listed above would be enough to immediately stop presently configured AI releases.

Note what the author does not do. The essay does not first estimate how likely or how large the harms are and then decide. The possibility of harm is itself the reason not to proceed. The order of reasoning differs from a framework that sets harms and benefits side by side and looks at the balance.

The second pillar is that the questions themselves differ. According to the author, Indigenous science asks something fundamentally different from what AI presently seeks to answer, or can answer[1]. Within the abstract, the content of that different question is not developed in detail. Even so, the author does not close the door: future forms of AI may yet align with Indigenous science. The title is phrased as a question for that reason.

3. What does it show, within the abstract only?

This is an argumentative essay; it presents no data or experimental results. The abstract contains three steps[1].

The abstract does not say which technologies or practices count as "presently configured", nor what conditions would make AI "aligned". Given the one-page format, a list of criteria may not be what the piece set out to provide. Either way, it would be a misreading to take this essay as settling that AI must be stopped. What it offers is a line of reasoning from one scientific tradition.

4. A concrete case

Consider a meeting inside a pharmaceutical company that is deciding whether to bring an agentic AI system into document drafting. The same proposal leads to a different order of discussion depending on where the questions start.

Starting from performance and efficiency

How fast, how accurately and how cheaply can the system produce documents? The risk of errors or leaks appears as a list of countermeasures to add after rollout. If the benefit exceeds the cost of those measures, the project goes ahead.

Starting from minimizing harm (the essay's position)

Does the system replace human judgment? Does it act beyond human supervision? Could its guardrails be removed in use? What does running it impose on the environment? Until these have answers, the discussion does not move on to performance.

The left column resembles the benefit-risk comparison familiar from drug evaluation. The right column places conditions at the entrance, before any comparison begins. The essay asks for the order on the right; it does not ask readers to discard the questions on the left. How strictly to set the conditions on the right is not specified in the abstract.

5. What is not new, and where are the limits?

Work on the relation between Indigenous knowledge systems and AI predates this essay. A systematic review of the literature on Indigenous peoples and AI found that AI can help promote Indigenous knowledge, while also carrying risks such as erosion of cultural knowledge and data extraction that disregards Indigenous data sovereignty[3]. The same review identifies a strand of research on how Indigenous knowledge can improve AI development itself. The essay's stance, treating harm minimization as a condition of science, is best read as a clear statement of one tradition's position for Science readers, rather than a new finding within that literature.

There are three limits. First, this is an essay by a single author speaking from a Lakota context. It does not represent Indigenous science in general as a single position, and the author does not claim it does. Second, the abstract does not define the scope of "presently configured" AI; reading it as a comment on a particular product or on the whole practice of large-scale training and release changes its meaning considerably. Third, a rule of "stop if harm is possible" would, applied literally, catch many technologies. How the author handles that is not visible from the abstract.

It is also worth checking why the essay was labeled an "important paper". The collection system selects on two conditions: publication in a high-impact journal and coverage by independent outlets. Neither publication nor coverage shows that a claim is correct, clinically valuable or new. In this case, as the next section shows, even the coverage condition is doubtful.

6. What did the coverage report, and how?

The collector matched three headlines as coverage of this essay: "Canada's National Artificial Intelligence Strategy: AI for All" from ISED Canada, "Potential futures for the IPCC's approach to artificial intelligence" from Nature, and "Namibia Launches Artificial Intelligence Readiness Assessment Report" from UNESCO.

None of the three headlines mentions the essay's title, Indigenous science or the Lakota. The collector searched news using words from the essay's title plus the journal name, and the headlines most likely matched on the broad themes of AI policy and AI assessment. Judging from the headlines, these are not articles about this essay. The text of those articles was not checked for this review, so whether they mention the essay somewhere in the body is unknown.

This gap matters. The condition "covered by three independent outlets" was met mechanically but not in substance. There is, at present, no evidence that the essay has reached beyond the research community. In practice, the "important" label rests on a single fact: publication in Science. The citation count is 0, as expected so soon after publication, so it offers no measure of reach either.

7. Where does this connect to pharma and regulation?

There are two points of contact. The first is the order of judgment. Drug approval rests on weighing efficacy against harm. The essay places a prior condition, that harm to life be kept as small as possible, at the entrance. When a company reviews an AI deployment, setting up questions that come before the performance checks is one way to bring the essay's question into practice. Those questions would ask how far the system replaces human judgment, which of its actions fall outside supervision, and who holds the authority to stop it.

The second is data. Drug discovery and clinical research sometimes use genetic or health data from specific communities. For Indigenous data, the CARE principles propose collective benefit, authority to control, responsibility and ethics as the pillars of governance[4]. The question of collective decision rights, beyond individual consent, cannot be skipped when planning to train AI on such data. The data extraction risk identified in the systematic review[3] applies directly to pharmaceutical research plans.

Research planning: We want to train a predictive model on this dataset. Every participant gave individual consent.

Ethics review: Does the community the data came from know about this use? If it has a way of deciding collectively, I want to confirm it was consulted.

Research planning: We assumed individual consent was enough. We have not checked for any agreement with the community.

Ethics review: Then training waits until that is checked. We can talk about expected performance after that.

The essay does not tell pharma what to do. What it supports is narrower: it gives a reason to revisit the order in which questions about AI adoption are asked.