Artificial Innovation

"Innovation is coming," I sang, "Innovation is striding in his suit of lights over moons and mountains, over parking lots and fountains, toward your silky side. Innovation is coming, he has a coat of many colors and all major credit cards and he is striding to meet you and culminate your foggy dreams in an explosion of blood and soil, at the end of the mechanical age."
- Donald Barthelme

The Innovation Factory is a global startup pitch competition run by the UN International Telecommunications Union to recognise AI-powered solutions that help address the Sustainable Development Goals. Throughout the year, startups apply to pitch at online and in-person events. A panel of judges selects the best pitch to win a free trip to the UN AI for Good Global Summit in Geneva for a chance to pitch at the Innovation Factory Grand Finale. Previous Grand Finalists include a blockchain-powered health service, the "world's first AI-native game engine for learning", and low-cost robotic limbs for amputees. (So a mixed bag, you could say.)

Despite running for nearly five years, the Innovation Factory has a low profile: the live pitch streams rarely crack 300 lifetime views and last year's Grand Finale was watched by less than a hundred people online. The real magic happens behind the scenes, of course, when finalists get the chance to rub shoulders with a "Multi-stakeholder, Interdisciplinary, Inter-generational" audience.¹ They sit down with ambassadors, take selfies with Gates Foundation program directors, and impress CEOs. They can forever claim the prestigious title of a UN-endorsed Good Innovation, for whatever value that may bring them in future business dealings.²

I first discovered the Innovation Factory in early May after learning that the competition had quietly passed through my city months ago, won by a local startup I'd never heard of. Pulling on that thread led me further into the Factory, and I've spent many sleepless nights trying to understand its strange machinery.

The AI for Good 2025 summit just kicked off this week with the ten³ qualifying startups pitching in the semi-finals on the 8th July, followed by the Grand Finale on the 9th. I've watched all of the qualifying pitches and read up on each company, and I have some thoughts.

Welcome to the Innovation Factory, please enjoy your stay.

The innovation assembly line

A note on ratings

I don't know the exact evaluation process employed by Innovation Factory judges, but my guess is that they probably would have been asked to rate each startup against a set of high-level criteria similar to this list, which includes areas like "Innovation", "Execution Strategy", "Technology & Risk Management", and "Global Relevance".

I think this is a pretty bad approach to rating startup pitches. We can broadly agree that startups need to do all of these things well to succeed⁴—but we're dealing with startups in their, shall we say, "pre-success" stage. A lot of these gaps can be filled in later; that's the point of the accelerator program. So I've come up with a streamlined rating system aligned with the competition's overall goal to "identify practical solutions using AI, scale those solutions for global impact, and advance the SDGs":

• AI: For better or worse, this is a competition for AI startups. I'm being very permissive about what counts as "AI" (good lord, I won't open that can of worms) but some form of AI needs to be an essential part of the product.
• Good: Does this solution solve world problems? Does it create any new ones?
• Innovation: Ideally, the startups should be doing something new and interesting.
• Factory: Is it plausible that this team can ship the product they're pitching? Are there any dealbreakers that could block them from ever shipping?

In the spirit of the competition, I've fed my ratings into an AI to derive a a numeric Artificial Innovation Score. Claude got really into the scoring, so I've left its comments in as well.

RevolutionAIze

RevolutionAIze is a Hyderabad-based startup working on MAAP, the Malnutrition Assessment & Action Plan. Their qualifying pitch was delivered by founder and CEO Romita Ghosh.

MAAP is a smartphone app for early detection of malnutrition in Indian children. It aims to solve the problem of error-prone manual height measurements by automatically measuring height from a single photo. The height data is combined with age and weight to detect stunted growth with "86% accuracy". The app also has nutrition planning and tracking features.

Ghosh first began work on MAAP in 2021 at a charity hackathon at IIT Hyderabad. She incorporated RevolutionAIze in 2022 with cofounder Nilashis Roy and the pair have continued work on the app since then, with support from various government and NGO initiatives.⁵ They're currently piloting the app with three Indian states and claim to have screened 100K children so far.

The MAAP website includes an AI-generated anime retelling of the company's journey.

When it comes to revenue, Ghosh takes a page out of the Gates Foundation handbook with the claim that "At heart, we believe profit and purpose can co-exist". I don't know if my heart shares that belief, but it certainly agrees that startups make more money when their R&D is heavily subsidised through public funding. RevolutionAIze estimates a Serviceable Available Market of $15 billion and plan to make money through subscriptions, partnerships, and algorithm licensing fees.

According to Ghosh, MAAP is built on two "patentable" proprietary ML algorithms and a proprietary dataset—presumably this is the data collected from Indian children in earlier trials.⁶ She explains that they needed to build "India-centric standards" because "the standard deviation for India is different" though it's not clear what this means. If the company has chosen to move away from the WHO growth standards and apply their own criteria for assessing malnutrition, it seems both reckless and unscientific to keep this commercially confidential.

What's more intereting to me is that there have actually been multiple attempts to build photo-based malnutrition detection apps for Indian children! Welthungerhilfe's Child Growth Monitor has been in development since 2018 and is fully open source. And just last year, researchers at IIT Hyderabad—the same institute that hosted Ghosh for the 2021 hackathon—piloted an AI-powered malnutrition assessment app with a local NGO. Maybe both of these are garbage and MAAP is much better, but the point is that the pitch doesn't mention them at all. If a founder is either unaware of similar solutions or unwilling to discuss them, it should raise serious questions about their credibiility.⁷

Viseur AI

Viseur AI is a platform for training and using AI models across radiology and pathology. I really don't know how to describe it better than this, so I'll defer to cofounder Serkan Sokmen's explanation from his pitch at the Türkiye qualifier.

We offer an on-promise service that allows all patient data to be examined with our platform, and we also provide [a] daily reporting and imaging process. [...] Our AI models learn from hospital data without moving it. Using federated and active learning strategies, each hospital trains the AIs and this keeps patient data safe, follows privacy laws, and helps the model get better with real world data. We apply this across all medical imaging devices such as CT, MRI, X-ray, digital pathology.
[...] Clinicians and developers can create their own AI-powered solutions by fine-tuning our state-of-the-art pre-trained models using their own labelled data, all without needing any coding or AI expertise. It's a true one-stop shop platform delivering everything they need.[^20]

Viseur was founded in mid-2024 and has raised €250K to date. They claim to have 11 existing customers—mostly hospitals—and have "achieved €200K in digital pathology sales, showing strong market adoption". (I'm not sure those numbers back up that claim, but OK.) The pitch implies that the platform is feature-complete, though it looks like they might not have finished building their radiology product offering because the radiology webpages are all "under construction".

One slide proudly states that "Trained AI models provide accurate and reliable results, up to 90%." A judge jumps on this during Q&A, pointing out that this isn't an acceptable threshold for digital pathology. Sokman admits that while the team comes from a radiology background, "we don't have any experience in digital pathology". This is a stunning admission from a company whose pathology software is apparently their only revenue stream.

Viseur AI has published no support documentation or video tutorials for clinical staff, no product demonstrations, and a scant handful of screenshots showing what the platform looks like in action. They have one technical document, the "Viseur AI PACS User Manual", but it's a little... odd.

Viseur AI is a HTML based package for PACS server which is designed to aid professionals in every day’s decision-making process, connecting all the medical data into a unified and fast performing network. Viseur AI ensures a fast and reliable way to search, present and analyze the medical data (images and video files) on various devices: computers, smart phones, tablets and so forth. Intuitive user interface, simple, but very powerful software controllable by touchscreen Medical Panel PC. Storing locally more than 125 hours of HD videos and up to 100.000 of still images.

The software in this manual bears little resemblance to the advanced AI platform described in the pitch. There's a reason for that—here's the user manual for the Softneta MedDream, a PACS product from a Lithuanian company:

MedDream is a HTML based package for PACS server which is designed to aid professionals in every day’s decision-making process, connecting all the medical data into a unified and fast performing network. MedDream ensures a fast and reliable way to [...]

I won't claim to know what's going on here. It's possible that Viseur has a white-label arrangement with the MedDream vendor, but then why wasn't this mentioned in the pitch? Where is the actual product that Viseur has built? Where is it??? How do you sell a product to a dozen hospitals without anyone asking to see proper support documentation?

Diawiser

This is a truly special pitch, and I'm really excited to tell you why.

Diawiser is building a smartphone app for voice-based glucose monitoring. Their qualifying pitch was presented by cofounder and COO Dr Sina Zare. In an interview with the ITU, Zare shares the emotional story behind the startup:

Diawiser started with a problem that hit close to home—literally. My partner [Diawiser cofounder & CEO Monika Fetingyte], who lives with type 1 diabetes, shared the daily frustration of managing glucose levels: the finger pricks, the constant monitoring, the unpredictability, and then the consequences – diabetic complications. It made me think: is there another way?
As a clinical pharmacist⁸ who’s always had one foot in healthcare and the other in AI innovation, I saw an opportunity. What if we could detect blood glucose abnormalities using just your voice? It felt radical at first, and maybe even a little crazy, but we dove in.

Until late last year though, the company was instead working on an therapeutic AI chatbot for diabetes management. When Zare was interviewed by a health tech accelerator in November 2023, he told a very different story about the company's origins:

My partner is a type 1 diabetic, and along with her we had to go through quite a few hassles when it came to disease management. Lack of information, limited access to healthcare professionals, and minimal personalized approach to disease management resulted in inaccurate decisions that lead to further complications that could have been avoided. My background in clinical pharmacy was of good help in making more accurate decisions, however, we decided to start a journey that would ensure all diabetic patients gain access to personalized clinically accurate actionable insights for their disease management.

Pivots are normal for startups, don't even worry about it.

The global diabetes market is huge. The average cost per month for testing supplies varies by region; many first-world countries subsidise this so that diabetics pay almost nothing out-of-pocket, but less well-off places like the United States, um, don't. The entire industry is an absolute racket and long overdue for some truly disruptive innovation. And Diawiser are really overturning the apple cart with a consumer-friendly freemium subscription model. Says Zare: "We believe in access ... imagine 10 million people using this for as low as 10 euros subscription."⁹ John Lennon would be sobbing.

There's no market competition for an app that can test blood glucose. This is a true revolution in diabetes testing. The alternatives are laughable: test strips are inconvenient, continuous glucose monitoring is too expensive, and "of course" Apple have been working on a glucose testing device—"like a watch or something"—but Zare waves this off as unimportant.¹⁰

When Q&A rolls around, everyone wants to talk about the tech. I've paraphrased the questions, but Zare's responses are shared in full because that's no less than he deserves.

What is the accuracy rate and how close are you to market?
Well, according to current scientific research, the accuracy is well within the range of 75 to 82%, depending on various factors that I don't think I'll have the time to say, but we can talk later about what impacts—there are challenges for the accuracy. And where we are, we are collecting data as if—um, basically, we are early stage and we are getting as many data as possible in order to improve our own accuracy. So if you're asking whether our product is ready for the users to be used, no. But we are getting there. We hope that our first pilot should be done by March, because we have a collaboration and we have a certain, like, requirements to be met within that timeline, as I mentioned, with Big Pharma and a couple of clinics. So that's—our first pilot will be ready by then.

Is there any independent validation that the tech works?
When it comes to voice exploration overall for all diseases, it's not highly explored yet. It's not oversaturated yet. And the data available with current research, for example from Mayo Clinic—one of the biggest ones—has done a small research on voice, and they have done it with glucose. So there is scientific evidence that this can work, but we are trying to, like, basically expand on it. It's not over.... explored.

How will your AI model account for the massive range of human voices?
First, you could look at it from a generalist perspective, meaning that you don't really need to train a generalist model that can detect anyone's glucose level coming in. You start with, you know, like Agile development. So what we're trying to do now is to train our small model with up to a certain number of data points, and with that, we can say that, OK, if it's—if you go with, um—trying to say that this person has high glucose or low glucose or normal, for this, you can do this with smaller number of people. And that's precisely what we're doing and that's precisely what we've done with a few diabetics that we know ourselves. As I mentioned, Monika herself is one, and we were able to detect that, OK, with really good accuracy, this is working. So the more number of people you put in, obviously, the better. We're not funded yet—heavily—so we are trying to show that this is possible to do. The larger data you put in, the better, obviously, the outcome. But also the model right now is smaller. If you want to go with larger, you need to have larger models.

What biomarkers are you using? How do you factor for precalibration of different devices? And how do you avoid false positives?
[At this point, the host chimes in to say there are only 30 seconds left for questions.]
Right. So if you want to talk about pre-calibration—so the concept of false positive would be that you will check it with your finger, it says you're high, you check, and then it's not high. You don't [??]—it's an additional method for non-invasive monitoring. So it helps people who don't have access to actually gain access. But let's talk later.

Yikes.

Zare's dismissive mention of the "small research" by the Mayo Clinic seems like a mangled reference to work by Klick Labs, a Canadian health R&D firm. In 2023, Klick published a paper in the Mayo Clinic Proceedings on using an AI model to detect type 2 diabetes through voice recordings. And in August 2024, Klick published another study showing a "small but significant" relationship between glucose levels and vocal fundamental frequency, although this comes with numerous caveats that the paper basically sums up with "more research needed."¹¹ So how did Diawiser find themselves in this cutting-edge research niche?

Zare and Fetingyte were hard at work pitching their diabetes chatbot throughout 2024, making it into the finals for awards like InnoStars and the Diabetes Center Berne Innovation Challenge. It's unclear exactly when they decided to refocus their efforts, but if I had to guess, I'd put my finger on sometime between the Klick Labs paper in August and the Innovation Factory Baltics application deadline in October.

It's stunning that Diawiser was awarded the top prize after that humiliating Q&A. The judges' voting took place behind closed doors and grew "quite heated", according to the host; reading between the lines I think at least one judge strongly resisted the nomination. I assume this was why the Chair of the Innovation Factory Baltics chapter took to the stage to give a rambling and insufferable talk about the importance of "connecting people—that's the Silicon Valley way".

I would eat my own fingers to know what was said in that room. Sadly, we can only speculate.

Project Ohm

Project Ohm aims to solve two pressing problems: AI's massive power consumption threatening global emissions targets, and Australia's "stranded" renewable energy that currently goes to waste and destabilizes the grid. Their solution is a decentralised network of AI computing nodes. An AI-powered orchestration platform will dynamically schedule AI training and inference workloads based on local energy av(ai)lability.

Behind this ambitious vision is Matthew Meszaros, a tech commercialization and business development veteran whose resume notably does not include any experience in energy infrastructure, decentralised computing, data centres, or artificial intelligence. Perhaps recognizing these gaps, Meszaros quickly assembled a team of three "Strategic Advisors" in the weeks leading up to his pitch at the Australian Innovation Factory in March.¹² As of early July, Project Ohm is on the hunt for an engineer to design and build the orchestration platform.

Project Ohm has not built any nodes yet, but they do have a node design and "two research papers that ... are patentable opportunities". Meszaros hopes to build the first PoC sometime around October on loaned hardware and a patch of dirt at a local university.

I believe this is a monumentally unviable business. I genuinely can't sum up all the red flags and technical issues into a coherent and readable narrative, so I'm just going to focus on one serious problem: They don't know what their product is.

One of the questions asked during the Q&A is "Who are your customers?" (a wild question for a business pitch to leave unanswered) and Meszaros explains:

So our two principal products in the, around the AI, is training [...] this is where we use epoch training and we use checkpoint, which matches our service. The second is big batch inferencing. So these two services are really important for enterprise customers who want to build trained models around their businesses. Then, you know, we've looked at fintech, who are looking at doing risk modeling [...] Then we're looking at biotech, who can use this for doing protein flowing[sic] or informatics, gene mapping, et cetera. There are a huge number of private research[ers] who are looking for an ESG-compliant solution and have government funding that requests that. So there's a huge, broad range — e-commerce who are looking at doing studies around their recommendations engines and so forth. So there's a huge, broad scope of customers that we can address with this.

But how exactly does Project Ohm plan to address those customers? The pitch entirely focuses on Ohm's competitive advantages against traditional data centres:

The key differentiating feature of Project Ohm is the AI-powered workload orchestration platform (which they haven't built yet, remember). But this is back-office stuff. Their potential customers have no reason to care about the clever infrastructure optimisation tricks happening behind the scenes. You know what those customers care about?

• Security: Governments and companies in regulated industries are required to use secure hosting facilities. A collection of shipping containers flung across university parking lots is unlikely to meet the stringent physical security requirements of an IRAP or SOC 2 certification.
• Availability and disaster recovery: What sort of uptime can customers expect, and what happens if a node gets knocked offline? How does the system handle data replication & failover for long-running jobs?
• Functionality: Other AI platforms include developer-friendly integrations, comprehensive observability features, cost management tools, one-click model deployments, single sign-on & RBAC, 24/7 technical support... none of these are mentioned in the pitch and don't seem to be a part of Project Ohm's plans.

That doesn't even get into the issue of cost. Existing providers like AWS and Microsoft benefit from enormous economies of scale. A highly distributed containerised computing provider is naturally going to have higher capital and operating expenses than a traditional data centre with the same compute capacity. What premium can customers expect to pay for this ESG-friendly barebones alternative to the mature products offered by competitors? Why would they pay that, when sustainability is already a massive selling point for every major cloud provider?

OptiCloud

For a radically different approach to sustainable cloud computing we turn to OptiCloud, pitched by cofounder and CEO Vijay Karia ("Technologist, Athlete, Planet Guardian"). Karia envisions a world where "every byte of data is used responsibly in support of the health of our planet", and to do this, OptiCloud is building an AI-powered platform that eliminates cloud waste through a suite of semi-autonomous & autonomous agents that perform cloud migrations, manage infrastructure, and update security configurations. The slide deck starts with AI-generated artwork and ends with a quote from Mahatma Gandhi.¹³

(If you're an SRE or security professional, please step away from the computer and take a moment to calm yourself.)

Their presentation includes a lot of numbers. Here are just a few:
- 50% of all digital resources are waste, resulting in $300B overspend and 75M tons of carbon emissions annually. OptiCloud could reduce those numbers by up to 60%.
- They're currently raising $50M with the goal of reducing all global carbon emissions by 1%. Investors can expect a 908% ROI over the first three years.
- Also, there are stretch goals! For every $500K in EBITDA, they will plant 10,000 trees in the Amazon; for every $1M, they'll fund the equivalent of 1K metric tons of CO2 reduction through renewable energy projects.
- At the end of the first year, they anticipate revenue of $20M; in three years, $840M.
- The team of 17 has a combined 20+ years of technical experience.¹⁴

Karia also tells us that the solution aligns with 12(!) of the 17 Sustainable Development Goals. There is some impressively creative accounting at work: for example, SDGs 5 and 10 ("Achieve gender equality and empower all women and girls" and "Reduce inequality within and among countries") will be achieved through re-investing in women-owned African startups. We're also told that "At the end of the day, you know, we're thinking to offer this service for free. At the very end of the day."

Until the sun sets, they've established a commercial partnership with a consulting firm who will market their platform to clients. Their scaling strategy is to form similar partnerships with other firms like McKinsey and KPMG.

There's not much to say about their tech because they haven't launched yet, but I think any competent engineer would be extremely alarmed by the concept of fully autonomous infrastructure management. Maybe their use of AI is very clever, but it's not magic: there are so many footguns in cloud configuration and IaC, and a lot of them are extremely hard to diagnose and correct without in-depth subject matter expertise.[^15]

[^15]
I'm just not convinced that OptiCloud's team has this expertise, given how Karia describes their IaC agent: "We can reverse engineer somebody's cloud in real time in Terraform. So, this used to take an engineer maybe 12 months to manually code these Terraform scripts; we've built a set of, you know, chain-of-thoughts—maybe a hundred different chain-of-thoughts—within our process that produce the real working Terraform scripts."
Bro, what? The difficult part isn't getting your existing resources into Terraform, it's developing robust, secure, and maintainable configs. This is where LLMs fail the hardest. Lots of people are working on AI-generated IaC and there are some promising results, but then that raises another question: why use OptiCloud's proprietary solution to generate IaC when many of these other solutions are developed in the open and can be tested and improved by the community?

I can't imagine a world where I'd sign off on this for any environments I look after. It's chilling to think about healthcare orgs, financial institutions, and governments hooking this up to environments with sensitive data.

To finish things off, here's a quote from Karia on the definition of " digital waste":

And you know, this is a big initiative. But if we collectively take action, and we clean up all our junk email, all our duplicate photos—I was just last week with NBC Universal and they have an archive of all these news videos that just sit there. And, you know, we're kind of adding more fuel to our problem as humanity. Because we're not taking out the trash.

Predictheon & Glidance

I've grouped these two startups together because they largely raised the same questions for me.

Predictheon is working on an AI-powered predictive monitoring solution for early detection of anesthesiology complications. The pitch is delivered by CCO Eva Gubern. The company was founded in 2019 by a team of clinicians and engineers, although work on the underlying models began around 2015. Predictheon raised an €840K seed round in 2021 and has also received a decent chunk of funding through programs like EIT Health. We're told that the company is two years ahead of its competitors, has strong interest from commercial partners, and is "Positioned for M&A". They're currently raising another €1.2M to complete the EU regulatory approvals process.

Glidance is building Glide, an AI self-guided mobility device for the blind, and is presented by founder & CEO Amos Miller. Glidance is by far the most successful startup in the Factory, with $2.4M+ pre-seed funding, $850K+ in preorders, and widespread media coverage. They're now raising seed to support initial delivery in 2025 and full commercialisation in 2026.

I don't know enough about the medtech industry to form a strong opinion on any of this, though nothing in the pitch struck me as particularly concerning. But I do have to question what Predictheon is doing in this startup competition in the first place.

Making surgeries safer is undeniably Good, but it is also very profitable: Predictheon has raised significant funding already and seems optimistic about their pathway to €28M net revenue in the next six years. Glidance has been massively successful and are already looking ahead to their Series A in 2026. How can you fairly compare them against a small community-focused startup struggling to raise their first $50K? Frankly, I just think all the other finalists have more to gain from the Innovation Factory, and that makes me wonder if Predictheon and Glidance displaced other startups that could have really used the leg up.

On a more uncomfortable note, consider that up to five billion people around the world have no access to basic surgical care at all. This technology doesn't help them; it primarily benefits individuals in wealthy nations where advanced medical care is already widely available. This isn't necessarily Predictheon's problem, but it certainly should be top of mind for the UN—SDG#3 literally says:

The United Nations is advancing universal health coverage and promoting equitable, people-centred health systems, focusing on fragile, conflict-affected regions and areas with significant health inequalities.

The Glide retails for $1500 with a $40 monthly subscription fee. The company is targeting North America and Europe for their launch, and one judge asks whether Miller has any plans for releasing to countries like Ethiopia (which has one of the highest blind populations in the world). Miller explains that Glidance are focusing on getting the tech right during their rollout, but that maybe in the future they'll be able to produce lower-cost devices for other markets. Once again, this is their prerogative, but there's a long history of companies releasing highly advanced assistive tech and abandoning it when it's not profitable, leaving early adopters out in the cold. The years of innovative award-winning research effort are locked behind patents and commercial IP so that nobody ever benefits from them again.

There's no such thing as trickle-down science. If the Innovation Factory organisers actually believe that AI startups can help advance the Sustainable Development Goals, they need to be thinking more critically about what sustainable innovation actually looks like, and whether the structures of VC-backed commercial tech startups are capable of delivering it.

Honourable mentions

These are the pitches that thought were genuine and interesting enough to deserve more serious consideration. I didn't spend as much time looking at these startups as the others, so I won't rate them.

Homai

Homai is "an AI-driven platform that helps to save Indigenous languages. By digitising endangered languages and integrating them into user-friendly devices and software, Homai enables cultural communities to keep their languages alive in daily use." I got the impression that Homai grew out of a direct community need and is very much a long-term passion project for the team. Last week, they shared a heartwarming video check-in about what the Innovation Factory win has meant for them in terms of creating new opportunities and connections.

Evet

Evet is building Agrocist, a vet advisory and recordkeeping tool for farmers. They've built their own AI model to help detect common poultry diseases from photos along with a chat interface (via Whatsapp); they also operate vaccine centers and a veterinary referral system. Their tech roadmap seems pragmatic, and importantly, they already have a track record of delivery. My only concern is that their mobile app seems to have disappeared in the last few months, possibly to be replaced by their Whatsapp chatbot. This feels like a questionable strategic move, but maybe it makes sense for them.

AI Diagnostics

AI Diagnostics has developed a digital stethoscope and custom AI model for early detection of tuberculosis. These guys are more of a question mark for me because the Africa pitching session was not streamed, but the tech sounds pretty neat and has already been through clinical studies. They're looking to develop more diagnostic models for pulmonary and cardiovascular diseases in the future.

Conclusion

Some snappy conclusion to tie everything together here!

Also link to this: https://www.techpolicy.press/the-ai-for-good-agenda-for-whose-benefit

How can conferences like AI4Good contribute? Consider the AI4Good Innovation Factory, which aims to "identify practical solutions using AI, scale those solutions for global impact, and advance the SDGs". We must abandon this very scale of thinking, which valorizes technological initiatives designed to grow quickly without needing to change. This mindset is emblematic of Silicon Valley innovation and disruption discourses.

Instead, AI4Good can fund projects for capacity building in local communities and non-profits, and critically engage with the limitations of how AI can potentially obscure alternative solutions. We should co-design bottom-up solutions with grassroots communities. The co-design process needs to start from jointly identifying and defining problems with affected communities, and considering all possible low-tech solutions without rushing to AI as the pre-defined solution. The design process needs to be led by local communities, not necessarily governments, large academic institutions or private companies, with support provided from traditional top-down power holders. Such initiatives can even help counteract harmful national agendas and holdgovernments or corporations to account—for example, the use of drones to produce reliable maps of the Borneo rainforest in Indonesia to to stop illegal deforestation driven by large-scale mining and palm oil plantations, or an activist-led effort in Mexico to create a detailed femicide map to challenge the government's lack of data transparency on the issue.