In response to Friday’s post on AI and Development, PS Baker asked when FP2P would publish its first AI generated post. Here you go.
My Question: Write a 600 word blogpost on the risks and opportunities of AI for international development, in the style of Duncan Green’s From Poverty to Power blog
ChatGPT’s response:
‘Introduction: Artificial Intelligence (AI) has emerged as a disruptive force with profound implications for international development. While AI presents exciting opportunities, it also carries inherent risks and challenges that need to be carefully navigated. In this blog post, inspired by Duncan Green’s thought-provoking style, we will explore the risks and opportunities of AI in the context of international development and discuss ways to harness its potential for the greater good.
- Opportunities of AI in International Development: a. Data Analysis and Decision-making: AI enables the processing of vast amounts of data, allowing for evidence-based decision-making. By analyzing data on health, agriculture, and other sectors, policymakers can gain insights that inform more targeted interventions and resource allocation, ultimately improving development outcomes.
b. Efficiency and Innovation: AI-driven automation can streamline administrative processes, optimize supply chains, and improve service delivery. This enhances the efficiency of development programs, enabling resources to be utilized more effectively. Additionally, AI fosters innovation by generating new solutions and approaches to long-standing challenges.
c. Addressing Inequality: AI has the potential to address inequalities by identifying and reaching vulnerable populations. Predictive analytics can help policymakers allocate resources more accurately, ensuring that marginalized communities receive the support they need. Moreover, AI can contribute to expanding access to education and healthcare, bridging the digital divide.
- Risks and Challenges of AI in International Development: a. Data Bias and Discrimination: AI systems are only as unbiased as the data they are trained on. Biased or incomplete datasets can perpetuate discrimination, reinforcing existing inequalities. It is vital to ensure that data used in AI models are diverse, inclusive, and representative of the populations they aim to serve.
b. Ethical Considerations: AI raises ethical dilemmas regarding privacy, consent, and the potential for misuse. Protecting individuals’ data and privacy rights while maximizing the benefits of AI requires robust regulatory frameworks. Ethical guidelines must be developed and implemented to ensure that AI is used responsibly, transparently, and for the common good.
c. Unequal Access and Technological Divide: AI adoption is not uniform globally, leading to a digital divide between countries and communities. Limited access to technology, inadequate infrastructure, and the absence of digital literacy can exacerbate existing disparities. Efforts should focus on reducing this divide, ensuring that AI benefits reach all sectors of society.
- Navigating the Path Forward: a. Emphasizing Human-Centered AI: AI systems should be designed with a human-centered approach, prioritizing human well-being and development objectives. This involves involving stakeholders in the design process and taking into account local knowledge, cultural sensitivities, and diverse perspectives.
b. Strengthening Governance and Accountability: Regulatory frameworks and ethical guidelines should be established to govern AI technologies in international development. These should promote transparency, accountability, and data protection while safeguarding against potential risks. Independent monitoring and evaluation mechanisms can ensure compliance and identify areas for improvement.
c. Capacity Building and Collaboration: Enhancing the understanding of AI among policymakers, practitioners, and communities is crucial. Capacity-building initiatives can empower individuals and institutions to leverage AI’s potential effectively. Encouraging collaboration among governments, civil society, academia, and the private sector can facilitate knowledge-sharing, resource pooling, and innovation.
d. Prioritizing Inclusivity: Efforts should be made to ensure that marginalized communities and countries are not left behind in the AI revolution. This requires investments in digital infrastructure, improved connectivity, and access to technology. Equipping individuals with digital skills and fostering inclusive AI ecosystems can mitigate the risk of exacerbating inequalities.
Conclusion: As AI becomes increasingly integrated into international development, stakeholders must navigate the risks and opportunities it presents. By seizing the potential of AI while addressing’
Erm, and there it stops (this is me now, promise). So first observation is the AI struggles with word limits. Other things I noted:
It struggles with style – I don’t think this sounds like me/FP2P (although impressed by its powers of flattery). More a standard academic presentation, but maybe a bit clearer.
The insights seem pretty bog standard (underwhelming)
And the writing! Clunky to say the least – ‘This involves involving’ – come on guys, that’s pretty bad.
Seems quite happy to exceed its brief. I never asked it for ‘ways to harness its potential for the greater good’, but it did it anyway. Lots of policy ‘shoulds’ that I did not request.
Very techno-regulatory – pretty feeble on politics and power – e.g. nothing about how AI is likely to interact with fragile/conflict affected/predatory states.
Overall, it seems more balanced on development implications than the UK government White Paper I wrote about last week, but that’s probably down to the framing of the question.
What else do you see?
It’s pretty meh isn’t it? I find that reassuring amongst all the talk of AI outsmarting humans.
Author
yep, pass rather than distinction, every time
Duncan, this is not meant as flattery, but I find you way more engaging and for sure funnier than AI 🙂 (and that’s coming from me (not AI).
Author
phew!
It does read a bit like a poorly prepared student with a basic knowledge of some key development areas trying to reach the word count in the completion of a late-night, last minute dot com, A level assignment. (Which is nothing like FP2P!) Though to be fair to AI the more detailed and explicit the AI prompt given, the better the answer. So it is worth persevering and trying to ask follow up questions, give more examples and guide the AI a bit more by giving feedback in order to craft a more powerful and useful article.
Author
be my guest Cliff, I think you’re right that the secret will lie in the skill of the prompt questions and follow ups, so why not see how smart you can get it?
It reads like a poorly-written conference communique. Or, possibly, a first year undergraduate trying to impress.
Did ChatGPT include that cartoon? That would be interesting!
Author
sadly no, that was me
Duncan – your job is safe (for now) 🙂 ….. but sounds like the bland generic stuff being churned out by strategy offices in most INGOs these days.
Author
Yep, looks like I won’t have to write any more posts like this….. ‘how to https://frompoverty.oxfam.org.uk/how-to-write-the-recommendations-to-a-report-on-almost-anything-introducing-friday-formulae/
Fantastic exercise. A decent enough summary to gather some points to chew on. Not a final product and definitely doesn’t have your personality and sense of humour, which is what makes reading your blog so engaging and personal – and I think it lacks perspective. It makes the same mistake as many policy organisations – thinking that EIDM is unbiased and neutral and will result in good advise. My question is “whose policy advice is AI reading?” If most sources are Northern centric, then AI will reproduce those thought patterns. So it is missing as a con. For good development decisions, we need to ensure more research from the southern actors is available for AI to train on. it is not just about ensuring the marginalised people have access to technology, it is about ensuring that marginalised countries are able to set agenda. At the recent On Think Tanks conference Think Tanks are starting to ask how do we get chatGPT to read our work? Great question for a blog?
Author
Great points Dena, do you fancy writing it?!
Hi, thanks for the invite. I think John Schwartz should write it – https://www.linkedin.com/in/soapbox/ – I will let him know 🙂
Indeed, today ChatGPT does not (yet) excel universally, and its primary feedback on any subject is average. Not surprising then when it comes to International Development, let alone in Duncan’s style…
But when thinking about the speed of learning these AIs are showing, the prospects are awesome, and I am glad that the G7 are committing to try some regulation of AI deployments.
Author
Thanks Mano, sounds like it’s going to go through the learning curve of School, College, Masters and PhD in double quick time and then eat all our lunches……
1. What version of ChatGPT did you use? And if it was GPT4, was it the web search enabled version, because this could have made a difference.
2. You get better results from most queries if you (a) spell out a few steps it can take to get to the end results (b) reiterate and revise your inquiry based on the initial result. Think dialogic.
Author
OK, now you’re exposing my ignorance Rick. I’ve subscribed to https://chat.openai.com/, but it doesn’t say what version it is or if it is web search enabled, as far as I can see. Definitely agree on rubbish in rubbish out tho
Use the ChatGPT+ version, then look at top of screen centre, then click on GPT4, then select browsing from the drop down menu that appears
Generated in what? a couple of minutes. Bland yes but fast and better than a lots of stuff I have seen generated over several days meetings! Even the confusing places such as the conflation of efficiency and innovation were food for thought. Good to know your job is safe. For now
Author
alas true Harriet!
This way of toying with AI was interesting months ago. By now, we should have stepped up discussions about its use and potential to a higher level, and using our brains. 🙂
Author
Feel free to lead by example!
Done, as many other people did! 🙂
Feel free to search for the rich debate.
Sitting in the CBA17 event for climate practitioners today we heard a lot about the challenges of tackling power and the realities of development processes reducing context specific insights into global lessons that don’t capture those realities. So i guess it is telling us with its very caveats (that it all depends on what it draws on) how much blah our global lesson learning reports contain.
Yes, this is what jumps out to me also. If AI draws on the existing writing and work product of our sector then this article says as much about the human thinking as it does the robotic summary. And so we see familiar themes coming out. Perhaps we need to be offering the bots with more original thought if we want to get a better result.
Try using GPT4 with the WebPilot plugin. I asked it to sumamrise the 5 most common themes in my 45 posts made between 2004 and 2010 on my “Rick on the Road” blog. here is its first result:
“Based on the analysis of the blog posts from 2004 to 2010, here are the five common themes ranked by their frequency:
1. Monitoring and Evaluation (M&E) in Development Aid Projects: This is the most common theme. The blog frequently discusses the importance of monitoring and evaluation in development aid projects, the challenges faced, and strategies for effective M&E.
2. Reflections on Development Aid: The blog often reflects on the state of development aid, including its successes, failures, and areas for improvement.
3. Case Studies and Experiences: The blog shares a lot of case studies and personal experiences from various development aid projects. These posts provide practical insights and lessons learned.
4. Tools and Techniques for M&E: The blog frequently discusses various tools and techniques for effective monitoring and evaluation. This includes both traditional methods and innovative approaches.
5. Public Transparency in Aid Processes: The blog often emphasizes the importance of transparency in aid processes. It discusses how transparency can be improved and the impact of transparency on the effectiveness of aid.
Please note that these themes are based on the content of the blog posts and the frequency of their occurrence. The ranking does not necessarily reflect the importance of these themes.”
You could try the same kind of analysis on your FP2P posts
fyi, rick
Author
Interesting, and can you ask it to do the same on stuff that’s not online? Wd love to see an analysis of the marksheets for my 60 students…..
Definitely prefer IRL Duncan.
I did something similar to write a blog using ChatGPT to address the fears of AI threatening communications roles. I did seven minutes of tidying/editing, and asked for comments from colleagues to see if they’d spot it… None did, although I suspect they were too nice to tell me the blog was incredibly bland.
You might be interested in a recent blog by my colleague, Sam Greene, who has been looking at ways local organisations can use AI to support their climate change adaptation work: https://www.iied.org/four-ways-local-organisations-can-use-ai-support-locally-led-adaptation-build-resilience
He’s just been at an event in Bangkok where he was surprised at how many climate practitioners and others were already using AI, although in a piecemeal way.
Very interesting, enriching discussion.
I definitely agree that GPT-4 cannot outsmart humans – yet. The technology will continuously improve as machine learning parameters are added, the underlying data broadens, and gets updated more regularly (the latter needs quite some computing power!). So, this is just the beginning, right?
From my experience, it is very true that clearer and more detailed instructions lead to better results. Recently I made GPT-4 create a qualitative self-assessment tool for a client based on the detailed household survey modules from the Self-Reliance Index (https://www.refugeeselfreliance.org/sri). I fed the survey questions and response categories to GPT-4 and prompted it to create such a tool. I had to prompt it a couple of times, refining my instructions until I got GPT to elaborate what I wanted: a 5-scale self-assessment tool ranging from 1 (lowest self-reliance) to 5 (highest self-reliance) which matches the SRI. I may do a quick write up of this soon. Of course, this tool is still in its infancy and needs to be refined and tested in case the organisation decides to use it. But it does show that GPT can very well complete complex tasks. I look at it as a first “point of departure” that will save humans substantial time.
An additional twist to this discussion:
In a short practice note I recently wrote (http://dx.doi.org/10.13140/RG.2.2.12099.78884), I turned the exercise around. Instead of asking GPT to summarise information (which I understand is very often done), I asked it to elaborate and interpret a set of aid principles. It is interesting to see that GPT substantially elaborated aid principles, extending the existing meaning, and giving very suitable ideas on how to measure and operationalise them. (As I did this in early 2023, I had used GPT-3.) This is of course not surprising given there are numerous such aid standards and tools to assess them available out there. Nevertheless, it does give a good idea of how GPT can be used to interpret and extend the meaning of existing text in the aid sector.