This is an excerpt from Blueprint2025, which will publish on January 15, 2025.
When I was writing Blueprint 2025 I started out trying an experiment to review the last 15 years. First, I commissioned an independent reader to review and let me know what they learned. Second, I used a large language model (LLM – the
methodology underpinning most of the Chatbots in use) to analyze 15 Blueprints
for me. I chose a tool called NotebookLM, a
Google product made specifically for writers to be a research and note-taking
tool that can quickly search, find text, and generate summaries.
Part of my reasoning was that Google already has sucked up all the
Blueprint content – it’s been on the web for years, is licensed as Creative
Commons, and I knew I wouldn’t be feeding the digital beast any information it
hadn’t already taken from me. Then I was then going to write a short piece
comparing what I learned from Susan and what I learned from NotebookLM.
I scrapped the 3-part idea. First, because I didn’t learn much from
NotebookLM. The few things I did learn are outlined below.
But I mostly scrapped the longer section because I don’t want to
encourage playing around like this with AI. I want to encourage you to be very,
very skeptical of how AI systems are being developed and by whom. I want you to
think twice, and then again, before playing with them with information from
your organization. I want you to read Jill Lepore’s words again, and think
about how it’s the slow drip, drip of promised convenience that embeds technology
in our lives in inextricable ways. I want to encourage you to seek out
noncommercial options and non-government options. I want you to be VERY clear
on the risks and benefits, to your mission, your constituents, and your
colleagues. And I want your organization to participate in building any such
systems in better ways – better for the environment, better for human rights,
better for your purpose.
Reflection 2: Text without Context
NotebookLM is a large language model (the same structure that
underlies chatbots such as ChatGPT) that uses a writer's own documents as its
source material. Google says it is designed to help writers gain insights
into their own documents faster. The team that developed it includes the
writer Steven
B. Johnson (The Ghost Map, The Invention of Air, and
other books).
To find out what insights NotebookLM could provide, I uploaded 15 Blueprints,
queried them in a variety of ways, and made a few observations on what I
learned in doing so.[i] An example of a question I
posed was “In what year did the Blueprint first discuss data donations?”
Through this and many other queries, I learned that the system is good at answering
questions that ask it to find facts within the pile of text, such as “what
year” or “how many.”
I also tried several of the pre-loaded questions that the system
prompts you to ask, such as “Create a thematic outline” of the documents. This
is basically what I asked Susan Joanis to do—read 15 Blueprints and tell
me what they argue. I learned that the AI uses certain types of text as
signposts—so subheadings and the tables of contents are transformed into
emphasis. Beyond that, nothing. NotebookLM
can only find text, it can’t add understand or add context, certainly not any
context beyond the words in its database. There’s a huge difference between
pattern-matching text (what AIs do) and understanding the context (what humans
do).
NotebookLM also provided misleading and false emphasis, which can best
be experienced through its Audio Overview. With the click of a button the site
will generate an audio summary – you can hear it here.
These fake voices attempt to add context by adding tone and emphasis; and,
indeed, they sound like real people. That’s frightening, precisely because it
sounds so real.
Here's what I learned:
this AI system is good at counting, pattern
finding, and it’s fast.
That's all.
These benefits come at a cost – I’ve paid it in my data, my IP, and my
time, all of which contribute to the growth of Google.
One of the last things I think we need right now is ways to manipulate
information. We already can’t tell truth from fiction. It’s not good for us as
readers, as neighbors, as professionals, or as citizens. It’s not good for
democracy to be creating and proliferating systems that further
corrode trust and truth. We’re already watching disaster
responders and city managers get attacked because of good old
fashioned human-spread
lies. Building systems that spread more lies, faster and further, is self-destructive. It’s
not good for disaster response; it’s not
good for democracy.
[i] I believe
the AI companies have illegally taken the copyrighted material of countless
authors and should be penalized. However, I wasn’t concerned about this for the
Blueprints as 1) I figured they live on the web, they’ve already been
used by every AI company; 2) I license them under creative commons to make them
easy to use; and 3) basically, everything in them is already publicly available
to do with as you (almost) please. There are a lot of other things that I would
not put into this system or any other AI.