A new study in the International Journal of Research and Innovation in Social Science surveyed 200 women micro-entrepreneurs in Bengaluru and reported that 32.5% of them are “actively using at least one AI-enabled tool” in their businesses. The tools, according to the paper, include KhataBook (a digital ledger), WhatsApp Business auto-replies, Canva, and Facebook’s “best time to post” suggestions.
None of those, as the women in the study experience them, are artificial intelligence.
The underlying research is not the problem. The Bengaluru paper makes a genuinely useful finding: women adopt digital tools when trusted NGO intermediaries provide sustained, peer-based training, and they abandon those tools when training is one-off or absent. That finding has been the consensus in ICT4D since Heeks established it in 2010.
What is new is the label. The paper describes a digital literacy intervention and calls it AI empowerment, and a peer-reviewed journal accepted that framing. This is not an isolated drafting error. It is the development sector’s pixie dust.
Call something, anything “artificial intelligence” and donor money rains down from the sky.
Subscribe Now to find AI pixie dust in your inbox!
The Inflation Is Now Routine
The Bengaluru study defines “AI tools” to include digital bookkeeping apps, automated WhatsApp replies, social media posting suggestions, and SMS inventory reminders. To a 2026 reader who has been promised generative AI, agentic systems, and machine learning models trained on billions of parameters, that list is a category error.
KhataBook does run machine learning on the backend for fraud detection and user profiling. None of that is what the women in the study are using.
- They are recording credits and debits in a digital ledger and sending payment reminders, a digital version of the paper notebook their grandmothers kept.
- WhatsApp Business pulls auto-replies from scripted templates.
- Canva’s posting-time suggestions are recommender heuristics that long predate the current AI wave.
Calling these “AI tools” inflates rule-based and template-based features into something they are not.
The authors are not unique in doing this. Tool vendors market the same conflation, grant applications copy the vendor language, and researchers writing for fast-turnaround journals adopt the grant language. The Bengaluru paper was received on 26 February 2026 and accepted on 3 March 2026, a six-day review cycle.
Whatever review took place at that pace was unlikely to interrogate definitional questions, and once a paper carries the “peer-reviewed” label, the next grant proposal can cite it as evidence that AI tools improve women’s business confidence by 73%.
SEC Fines for AI Washing
In March 2024, the SEC settled charges against Delphia and Global Predictions for telling investors they used AI and machine learning when they did not. The two firms paid $400,000 in combined civil penalties. SEC Chair Gary Gensler used the term “AI washing” and stated that “investment advisers should not mislead the public by saying they are using an AI model when they are not.”
Academic researchers have since defined AI washing as the practice of exaggerating, misrepresenting, or falsely claiming AI capabilities to gain reputational or financial advantage.
Donors are not the SEC. No Form ADV exists for a development implementer who promised AI-driven decision support and delivered a templated chatbot. The financial sector got fined into precision. The development sector has no equivalent forcing function, which is why the inflation is accelerating.
The Cost Is Not Rhetorical
When the next funding cycle asks for “AI for women’s empowerment” programs, three things will happen.
1. The same work will be repackaged with a new label.
This is the lowest-cost path for implementers who already know the model works, and it is the path the Bengaluru paper makes easier.
A donor reading “73% of women using AI tools with NGO support reported greater confidence in decision-making” has no way to know that the underlying tool is a digital ledger. The donor funds AI. The implementer delivers digital literacy.
Both parties get what they wanted and the evidence base on actual AI tools never gets built.
2. AI tools will get deployed on top of this inflated evidence.
When a credit scoring vendor, a generative chatbot for health advice, or an AI-driven fraud detection system enters the same market, they will cite “established evidence” that AI tools empower women micro-entrepreneurs in LMICs.
The cited evidence will be studies like the Bengaluru paper, which tested no such thing. We saw the same dynamic with mobile money, where early evidence of basic SMS payments was used to justify deployment of complex algorithmic lending products that produced very different outcomes.
3. The real finding gets lost.
The Bengaluru paper’s real contribution is that 35% of non-adopters cite fear and lack of confidence as their primary barrier, not cost or infrastructure. That is a finding worth building on, because it tells donors that handing out smartphones and subsidizing data plans does not address the binding constraint.
But that finding is buried under the “AI” framing, which directs reader attention toward the technology and away from the social infrastructure that the paper documents working.
Who Has Responsibility
If “AI” can describe a digital ledger and a Facebook hashtag suggester, the term has no meaning in development research and we should stop using it.
Yet machine learning systems are entering the sector now, in health decision support, credit scoring, and humanitarian logistics, and we need a working term to describe them and the specific risks they create.
But the responsibility for keeping the term meaningful is distributed, and right now nobody pushes back.
1. Researchers should define their terms.
A paper that calls KhataBook an AI tool needs to either justify that claim against a specific technical definition or use a more accurate label. “Digital tools,” “mobile business apps,” and “smartphone-based business aids” are all available and all more accurate. Peer reviewers should send back manuscripts that conflate the categories.
2. Journals should slow down.
A six-day peer review cycle is incompatible with catching definitional drift on a fast-moving technology. Editors at IJRISS and similar venues should require authors to specify the model class and technical mechanism for any tool described as AI. This is what SEC enforcement standards already require of regulated entities. Academic publishing should not be the soft target.
3. Donors should require specification.
Every grant proposal using “AI” should state the model class, training data provenance, and decision-making locus. If the proposed intervention turns out to be rule-based template messaging, fine, fund it as that. If it is a fine-tuned language model giving health advice to community health workers, also fine, fund it as that. But the donor should know which one they are buying.
4. Implementers should resist the marketing.
When a vendor offers an “AI-powered” version of a tool you already use, ask what specifically is AI about it. If the answer is recommender heuristics that predate transformers, you can write your proposal more honestly and your evaluation will be more useful.
The Broader Pattern
There is a vast gap between what women micro-entrepreneurs need and what donor programs deliver. The pattern is consistent. The sector reaches for whatever technology is currently fundable, attaches it to a population that needs sustained support of a different kind, and produces evidence that the technology worked when what worked was the support.
The Bengaluru paper is honest about this in its own findings.
The strongest predictor of confidence in the study is not the tool, it is the training hours. Women who got more than five hours of NGO support reported confidence scores between 7 and 9. Women who got less than two hours scored between 2 and 4. The technology is the constant; the variable that explains the outcome is the human relationship.
As the paper says, AI tools alone do not empower, people do.
We have known that for fifteen years. We did not need new technology to learn it. What we are doing now is using new technology vocabulary to recycle old findings into new funding cycles, and the cost is that we are losing the ability to distinguish between rebranded digital literacy and real algorithmic intervention.
When algorithmic interventions go wrong, and they will, we will not have the conceptual vocabulary or the evidence base to investigate them, because we burned both rebranding ledger apps. The SEC has shown us the path. Researchers, journals, donors, and implementers in the development sector should not wait for an equivalent enforcement mechanism to develop the discipline.
Now Read These Related Posts