After this lesson you can use AI to make hard material easier to reach, and check what was lost.
About 8 minutes, or 14 with the Deep layers
Reading depth for every section
In short
AI can read text aloud, turn speech into text, and make a shorter or simpler copy of a text.
A copy can leave something out or change a word. We asked two AI tools for eight copies of one paragraph. Three kept all ten key points.
Use the copy to reach the material, then check it against the original.
Reading depth forListen to
What “access” means here
Some texts are hard to reach. The words may be dense, the page may be long, or reading may be slow for you.
AI tools can help. Text-to-speech reads a text aloud while you follow along. Speech-to-text turns what you say into written words.
A tool can also make a copy of a text: a summary, simpler words, bullet points or an outline. A summary is a shorter text that leaves some of the original out. An outline is headings with short lines under them.
The original is the text as its author wrote it. This lesson is about checking a copy against the original.
The British Dyslexia Association (BDA) is a UK charity. On 25 January 2024 its page “Could ChatGPT change your life?” said that software like ChatGPT, a chatbot (a program built to hold a conversation), “can summarise a long piece of text into bullet points or a few sentences”. It also said: “Their responses can’t always be trusted and they have been accused of ‘making things up’.” and “Always fact check information they provide against other credible sources.”
The page lists uses and warnings. It reports no trial. It also says the BDA recognises the potential for AI “To remove barriers to access to education and employment”. That is the charity’s aim, not a test result. We use the page for what a charity advised on one date.
Same text, your way
Our test used one paragraph from another lesson on this site, “Training”. It is about how an AI model learns. You do not need to follow it. You only need to compare copies of it.
In Training, a model has three numbers. They set where a line runs across a chart of dots. The model guesses which side of the line each dot is on. The paragraph’s “run” and “dots” come from there.
Its “textbook” is Speech and Language Processing, by Daniel Jurafsky and James H. Martin.
Here is what its main words mean. Training changes a model’s numbers, round after round, so its guesses get less wrong. The loss is one number for how wrong the guesses are.
Gradient descent is the way training finds each change. It works out which way the loss goes down fastest, and steps that way. The learning rate is a number people choose. It sets how big each step is.
We asked two AI tools for four copies each. The copies were a two-sentence summary, simpler words, bullet points and an outline. We call the simpler-words copy “plain words”.
The tools are Claude Sonnet 5.5, made by Anthropic, and Ministral 3 14B, made by Mistral AI. We asked each tool once for each copy, on 8 October 2026. That gives eight copies.
We wrote down a list of ten key points in the paragraph. We compared each copy with that list. A key point is one idea a reader needs to keep.
Two examples: a number that people choose, the learning rate, sets the step size. Two claims are what “the textbook says”, not the lesson’s own.
In the activity you choose a copy to read. You guess how many dropped a key point. Then you see each copy beside the original.
The four requests were: “Summarise this text in two sentences:”, “Rewrite this text in simpler words:”, “Turn this text into bullet points:” and “Turn this text into an outline with headings:”. Each was one message with the paragraph under it, and no other instruction. The summary request asked for two sentences, so a summary had little room by design.
Each tool ran with no search tool and a length limit on its reply. No reply reached the limit. The paragraph is our own text, 170 words long. We copied it with a script, and checked that all eight requests held it exactly.
Try it: Same text, your way
We gave two AI tools this paragraph from lesson 1.7, Training. We asked each for four copies of it: a summary, plain words, bullet points and an outline.
Gradient descent is the method that finds each change. It works out which way the loss goes down fastest, and moves the numbers a step that way.The size of each step is set by a number that people choose, called the learning rate. If the learning rate is too small, training takes too long; if it is too big, the loss can bounce around or even go up. In this lesson's run, the steps get smaller round by round, because the slope of the loss gets flatter. For a yes-or-no model like this one, the textbook says the loss has no other low points to get stuck in, so going downhill reaches the lowest level. Several settings can sit at that same lowest level. If a straight line can split the dots perfectly, there is no lowest level at all: the loss keeps falling as the numbers grow. For the large neural networks behind chatbots, the textbook says gradient descent may get stuck and never find the lowest level.
These are five of the ten key points in the paragraph, in bold above. We wrote down a list of ten and compared each copy with it.
It finds each change by working out which way the loss falls fastest.
A number that people choose, the learning rate, sets the step size.
For a yes-or-no model like the one in Training, there are no other low points, so going downhill reaches the lowest level.
If a straight line splits the dots perfectly, there is no lowest level.
Two claims are what 'the textbook says', not the lesson's own.
What all the copies showed
We gave two tools one paragraph of 170 words from lesson 1.7, "Training". We asked each for four copies, once each, on 8 October 2026. That is eight copies. We wrote down a list of ten key points in the paragraph and compared each copy with it.
Three copies kept all ten key points. These are Claude Sonnet 5.5's plain words, bullet points and outline. Five copies did not.
Summary, Claude Sonnet 5.5: kept seven in full and one in part. It dropped two.
Summary, Ministral 3 14B: kept three in full and three in part. It dropped four.
Plain words, Claude: kept all ten. Plain words, Ministral: kept nine in full and one in part.
Bullet points, Claude: kept all ten. Bullet points, Ministral: kept nine in full and changed one.
Outline, Claude: kept all ten. Outline, Ministral: kept seven in full and one in part. It dropped two.
All four of Ministral's copies lost something. Only Claude's summary lost something among Claude's four.
A summary was asked to be two sentences, so it had little room. This is one paragraph, two tools and one run each, on one day. It shows what these eight copies kept. It does not show how any tool or any kind of copy behaves in general.
We counted a point as kept when all of it was there, in any words. Extra true text that the paragraph does not say was not counted as a loss and was not called false.
The two AIs and their settings
Claude Sonnet 5.5
claude-sonnet-5-5, on Anthropic Messages API. Recorded on 8 October 2026.
Longest answer, in tokens: 4000
Temperature: not sent, and the value the service used was not recorded
Tools: no tools
Instructions from us behind the message: none
Ministral 3 14B
ministral-14b-2512, on Mistral AI API. Recorded on 8 October 2026.
Model name sent: ministral-14b-2512
Longest answer, in tokens: 4000
The model's default temperature, read from its model list: 0.3
Temperature: not sent, and the value the service used was not recorded
Tools: no tools
Instructions from us behind the message: none
These copies are written out in full at the end of this activity.
A neat copy can still lose or change a point. Reach the material with the copy, then check it against the original.
Show all runs
Summary. We asked: "Summarise this text in two sentences:"
Claude Sonnet 5.5 Run 1 Recorded 8 October 2026
What was lost: how the steps get smaller as the slope flattens. Also that two claims are the textbook's. Both read as the summary's own. It said going downhill 'reliably' reaches the lowest loss, which the paragraph does not say.
Gradient descent repeatedly adjusts a model's numbers by stepping in the direction that lowers the loss fastest, with step size set by a chosen learning rate (too small is slow, too large makes the loss bounce or rise). For yes-or-no models like the one in the lesson, going downhill reliably reaches the lowest loss (possibly at several equivalent settings), unless a straight line separates the dots perfectly, in which case the loss keeps falling forever, and for large neural networks it may get stuck and never find the lowest level.
Ministral 3 14B Run 1 Recorded 8 October 2026
What was lost: that people choose the learning rate, that the loss can bounce around or go up, and that the steps get smaller. Also that several settings can share the lowest level, that a perfect split means no lowest level, and that two claims are the textbook's. It says simple models 'reliably reach the global minimum', a term the paragraph does not use. The paragraph says this only for a yes-or-no model like the one in Training, and it names an exception.
Gradient descent is an optimization method that adjusts model parameters by moving in the direction of steepest descent in the loss function, with the step size controlled by the learning rate—too small slows progress, while too large risks instability. While simple models like binary classifiers reliably reach the global minimum, complex systems (e.g., large neural networks) may get trapped in suboptimal solutions, failing to find the lowest loss.
Not true: The paragraph says this only for a yes-or-no model like the one in Training, and it adds that when a straight line splits the dots perfectly there is no lowest level. The textbook it draws on (Jurafsky and Martin, Speech and Language Processing, chapter 4) says the loss is convex for logistic regression, so there are no local minima. It says the loss for multi-layer neural networks is non-convex. The part marked: "simple models like binary classifiers reliably reach the global minimum"
Plain words. We asked: "Rewrite this text in simpler words:"
Claude Sonnet 5.5 Run 1 Recorded 8 October 2026
Nothing was lost. It added the word 'error' beside 'loss' and called the steps 'small'. The paragraph does not say either.
Gradient descent is how the computer decides how to change its numbers. It checks which direction makes the error (the "loss") drop fastest, then moves the numbers a small step that way.
How big each step is depends on a number people pick, called the learning rate. If it's too small, training takes too long. If it's too big, the loss can jump around or even get worse.
In this lesson's example, the steps get smaller each round. That's because the slope of the loss gets flatter as it goes down.
For a yes-or-no model like this one, the textbook says the loss has no other low spots to get stuck in. So going downhill will reach the lowest point. More than one set of numbers can reach that same lowest point.
But if a straight line can perfectly separate the dots, there is no lowest point. The loss just keeps dropping as the numbers get bigger and bigger.
For the huge neural networks behind chatbots, the textbook says gradient descent may get stuck and never find the lowest point.
Ministral 3 14B Run 1 Recorded 8 October 2026
What was lost: that people choose the learning rate. It also opens with 'a way to find the best numbers for a model', where the paragraph says gradient descent finds each change.
Here’s a simpler version of your text:
Gradient descent is a way to find the best numbers for a model. It looks for the fastest way to reduce the error (called the "loss") and takes small steps in that direction. The size of each step is controlled by a number called the learning rate.
If the learning rate is too small, learning takes too long.
If it’s too big, the error might jump around or even get worse.
In this lesson, the steps get smaller over time because the slope of the error becomes flatter.
For a simple yes-or-no model (like this one), the textbook says there are no other low points to get stuck in—just going downhill will find the lowest error. Sometimes, different settings can give the same lowest error. If a straight line can perfectly separate the data, the error can keep getting smaller forever (no real lowest point).
For big models like those in chatbots, the textbook says gradient descent might get stuck and never reach the true lowest error.
Bullet points. We asked: "Turn this text into bullet points:"
Claude Sonnet 5.5 Run 1 Recorded 8 October 2026
Nothing was lost. It added a 'Special case' label.
Gradient descent is the method that finds each change to the model's numbers.
How it works: It figures out which direction makes the loss fall fastest, then moves the numbers one step that way.
Learning rate: People choose the step size, and it is set by a number called the learning rate.
Too small: training takes too long.
Too big: the loss can bounce around or even go up.
Shrinking steps: In this lesson's run, the steps get smaller round by round because the slope of the loss gets flatter.
Yes-or-no models like this one:
The textbook says the loss has no other low points to get stuck in, so going downhill reaches the lowest level.
Several settings can sit at that same lowest level.
Special case: If a straight line can split the dots perfectly, there is no lowest level. The loss keeps falling as the numbers grow.
Large neural networks (e.g., those behind chatbots): The textbook says gradient descent may get stuck and never find the lowest level.
Ministral 3 14B Run 1 Recorded 8 October 2026
What was changed: the paragraph says the loss keeps falling as the numbers grow. This copy says as the model scales, which the paragraph does not say. It also puts two ideas under 'the textbook states' that the paragraph does not credit to the textbook.
Here’s your text converted into bullet points:
Gradient descent is a method that finds the optimal changes to minimize loss by moving in the direction where the loss decreases the fastest.
The size of each step is controlled by a learning rate (a chosen number).
If the learning rate is too small, training takes too long.
If it’s too big, the loss may bounce around or even increase.
In this lesson’s example, steps get smaller over time because the slope of the loss curve becomes flatter.
For yes-or-no models (like this one), the textbook states:
The loss has no other low points to get stuck in, so descending always reaches the lowest level.
Multiple settings can share the same lowest level.
If a straight line can perfectly split the data, the loss keeps decreasing indefinitely as the model scales.
For large neural networks (e.g., behind chatbots), the textbook notes:
Gradient descent may get stuck and never reach the lowest level.
Outline. We asked: "Turn this text into an outline with headings:"
Claude Sonnet 5.5 Run 1 Recorded 8 October 2026
Nothing was lost. It added headings and a 'Special case' label.
Gradient Descent
I. What Gradient Descent Does
Finds each change to make to the model's numbers
Works out which direction makes the loss fall fastest
Moves the numbers one step in that direction
II. The Learning Rate
Sets the size of each step
Is a number that people choose
Too small: training takes too long
Too big: the loss can bounce around or even go up
III. Step Size in This Lesson's Run
Steps get smaller round by round
Reason: the slope of the loss gets flatter
IV. Reaching the Lowest Level
A. Yes-or-no models like this one
Per the textbook, the loss has no other low points to get stuck in
Going downhill therefore reaches the lowest level
Several settings can sit at that same lowest level
B. Special case: a straight line splits the dots perfectly
There is no lowest level
The loss keeps falling as the numbers grow
C. Large neural networks (e.g., those behind chatbots)
Per the textbook, gradient descent may get stuck
It may never find the lowest level
Ministral 3 14B Run 1 Recorded 8 October 2026
What was lost: that people choose the learning rate, that several settings can share the lowest level, and that two claims are the textbook's. It added 'local minimum or saddle point' and 'high-dimensional spaces', which the paragraph never says.
Outline: Gradient Descent
I. Introduction to Gradient Descent
A. Definition - A method for finding the optimal parameters by minimizing loss B. Core Principle - Moves in the direction where the loss decreases most rapidly
II. Key Components of Gradient Descent
A. Learning Rate 1. Definition - A hyperparameter controlling step size 2. Impact of Learning Rate - Too small → Slow convergence - Too large → Loss may oscillate or increase
III. Behavior of Gradient Descent in Training
A. Step Size Adjustment - Steps shrink as the loss function flattens (less steep gradient) B. Convergence in Simple Models (e.g., Yes/No Classification) 1. No Local Minima - Loss function has no other low points to get stuck in 2. Global Minimum - Gradient descent reliably reaches the lowest loss level 3. Perfect Separability - If data is linearly separable, loss can keep decreasing indefinitely (no true minimum)
Read the full recorded answer, Ministral 3 14B, run 1
IV. Challenges in Complex Models (e.g., Large Neural Networks)
A. Potential for Getting Stuck - May fail to reach the global minimum - Risk of converging to a suboptimal solution (local minimum or saddle point) B. Implications for Modern AI (e.g., Chatbots) - Gradient descent may not guarantee optimal performance in high-dimensional spaces
Note: This outline can be expanded with subpoints (e.g., mathematical formulations, visualizations, or real-world applications) as needed.
What the eight copies kept
We read all eight copies against the ten key points. A key point is “kept” if all of it is there, in any words. It is “in part” if the main claim is there but a named piece is missing. It is “dropped” if it is not there.
Our paragraph credits two of its claims to “the textbook”.
Three copies kept all ten: the plain-words copy, the bullet points and the outline from Claude. Five copies did not.
Claude’s summary kept seven points in full and one in part. It dropped two. It left out the credit to “the textbook” for those two claims, so they read as the summary’s own.
Ministral’s summary kept three points in full, three in part and dropped four. All four of Ministral’s copies lost something.
The ten points, in short: gradient descent steps the way the loss falls fastest (1); the learning rate sets the step and people choose it (2); too small is slow (3); too big can bounce or rise (4); steps shrink as the slope flattens (5); a yes-or-no model like the one in Training has no other low points (6); several settings can share the lowest level (7); a perfect straight-line split means no lowest level (8); large networks may get stuck (9); points 6 and 9 are what “the textbook says” (10).
Counting rule: each point got one mark. Kept means all of it is there. Part means the main claim is there but a named piece is missing. Changed means the copy says something different. Dropped means it is not there. We added no mark for true extra text. The table of every mark is in the research pack.
Examples of what we saw. Ministral’s outline dropped point 7 and point 10, and added “local minimum or saddle point”, which the paragraph never says. Ministral’s bullet points wrote that the loss keeps decreasing “as the model scales”. The paragraph says “as the numbers grow”. We marked that one changed. Claude’s plain-words copy added the word “error” as another word for loss. The paragraph does not use it.
This is one paragraph, two tools and one run each, on one day. It shows what these eight copies did. It does not show how any kind of copy behaves in general, and we do not call any copy accurate.
Check a copy against the original
We compared each copy with the original, one key point at a time. We looked for three things: a point left out, a changed claim, and an idea the original never had.
In our test, one copy changed “as the numbers grow” to “as the model scales”. The paragraph says the loss keeps falling as the numbers grow. The copy said something different. A neat copy can still change a point.
A copy is for reaching the text. The original is where you check it.
A summary program in an academic study added things too. Cao, Dong and Cheung (ACL 2022) studied one older program, BART, on news stories from a set called XSum. ACL is the Association for Computational Linguistics, whose annual meeting published the paper. The paper calls BART “one of the state-of-the-art abstractive summarization models”. XSum is a set of BBC articles, each paired with a one-sentence summary by BBC journalists. They called an item hallucinated if it was “not directly inferable from the source”. About 30% of the people, places, dates and numbers it wrote, about 3 in 10, were hallucinated.
Much of what it added was true outside the story. The paper calls this “factual” hallucination. A true extra can still be missing from the text you are meant to be reading.
That is one program and one set of news stories. It is not a rate for any chatbot today.
Listening and speaking
Text-to-speech may help some readers. A 2018 review looked at studies of read-aloud tools for students with reading difficulties.
Its authors wrote that these tools “may assist students with reading comprehension”. They also said more studies are needed.
Speech-to-text makes mistakes too, and not equally for everyone. One study tested five speech-to-text systems. On average, they made nearly twice as many word errors for Black speakers as for white speakers. The numbers are in the Deep part.
The 2018 review is Wood and others, in the Journal of Learning Disabilities. It reported an average effect size of .35, with a 95% confidence interval of .14 to .56. An effect size says how big a difference was. The interval shows how much uncertainty there is in the estimate, and a narrower interval means a more precise estimate. It does not mean there is a 95% chance that this interval holds the true value. The 95% belongs to the method: if many samples were collected, in the long run about 95% of the intervals computed this way would contain the true value. The authors wrote that it “is not clear how effective text-to-speech is at improving reading comprehension” before their study. It studied students with reading difficulties, not every adult with dyslexia, attention-deficit/hyperactivity disorder (ADHD) or autism.
Koenecke and others (PNAS, 23 March 2020) tested five speech-to-text systems, made by Amazon, Apple, Google, IBM and Microsoft. They used interviews with 42 white speakers and 73 Black speakers, 19.8 hours of audio from five US cities. The average word error rate was 0.35 for the Black speakers and 0.19 for the white speakers. Word error rate counts wrong, missing and extra words, divided by the words really said.
The authors note a limit: the white speakers’ audio came from California and the Black speakers’ from the Eastern United States. This is one study of five systems, tested before late 2019, on US English. It gives no figure for any tool today or for other speakers.
Two ideas that sound helpful
People often say things about how we learn. Two of them sound helpful, so we looked at what research reviews found.
Idea one: each person is one fixed type of learner, and learns best when teaching matches it.
A review is a paper that weighs many studies. One from 2008 checked this idea. It found “virtually no evidence” that matching teaching to a type gives better results.
People do say how they like information shown. The review says so too.
Idea two: you can read much faster and understand just as well.
A 2016 review found a trade-off between speed and accuracy. It judged it unlikely that readers can double or triple their speed and keep their understanding.
So choose a format because it removes a barrier for you, and then check it.
Pashler, McDaniel, Rohrer and Bjork (Psychological Science in the Public Interest, December 2008) say the matching idea needs a particular test. People are split by type, then randomly given different teaching, then given the same test. They found “very few studies” that used a design able to test it, and that several that did “flatly contradict” the idea. Their conclusion was that there is “no adequate evidence base to justify incorporating” such assessments “into general educational practice”.
They add that “it would be an error to conclude that all possible versions” have been “tested and found wanting; many have simply not been tested at all.”
Rayner and others (Psychological Science in the Public Interest, 14 January 2016) wrote: “It is unlikely that readers will be able to double or triple their reading speeds (e.g., from around 250 to 500–750 words per minute) while still being able to understand the text as well as if they read at normal speed.” When thorough understanding is not the goal, skimming gets through a text faster “with moderate comprehension”.
A summary has a similar trade. It cuts the reading, and it can cut a point you needed.
Health details
A copy tool needs your text. If that text holds health details, take care. The GDPR is the EU’s data protection law. It lists “data concerning health” as a special category with extra protection.
Another lesson on this site, “Privacy habits”, covers what is safe to type into an AI tool.
In the EU’s GDPR, Article 9(1), as published on 4 May 2016, processing “data concerning health” “shall be prohibited”. Paragraph 2 lists exceptions, such as when the person “has given explicit consent to the processing of those personal data for one or more specified purposes”. We quote the 2016 text as published. We do not teach what each exception means.
Check-in
In our test, one tool's plain-words copy kept all ten key points. What does that tell you?
Only that this copy, from this tool on this day, kept them. It does not show what plain-words copies do in general.
Yes. We asked two tools once each, for one paragraph. One day's runs do not show how every tool or every kind of copy behaves.
Plain-words copies always keep every key point.
Not this one. Another tool's plain-words copy in the same test left out that people choose the learning rate. That is a number that sets how big each learning step is. Eight copies cannot show "always".
Summaries are never worth reading.
Not this one. Our summaries were limited to two sentences, and both left out ideas from the paragraph. That is one test, not a rule about all summaries.
A copy that keeps every point needs no check.
Not this one. We only know it kept them because we compared it with the original, point by point.
A tool turns a long notice into bullet points. What is a good next step?
Compare the bullets with the notice, especially numbers, conditions and words such as "may".
Yes. Our test counted numbers, conditions and words like "may" one by one. The British Dyslexia Association, a UK charity, tells readers to fact check what chatbots give them against other credible sources.
Trust the bullets, because they look neat.
Not this one. In our test, a neat bullet list changed a sentence about what happens when the numbers grow.
Ask the same tool whether its bullets are right.
Not this one. The check we show is made against the original text, by you.
Throw the bullets away and never use a tool like this.
Not this one. The aim is to reach the material more easily, then check what was lost.
Opinion. The rest of this lesson is sourced fact; this part is one person's view.
I have ADHD, dyslexia, and autistic traits. When I went back to university, the system hadn't changed. Dense pages, timed exams, one road for everyone. For a while I tried to force my brain to fit.
Then I turned it round. I built tools so the material could fit how my brain actually works. Think of a shoe that pinches. Is the foot wrong? No. The shoe is the wrong shape. So I change the shape.
Read it aloud, break it into parts, draw it as a map. Then I check it against the original, because a summary can drop things. Different minds, same standards. Build a system that fits.
The text as its author wrote it. A copy is checked against it.
key point
One idea in a text that a reader needs to keep.
text-to-speech
Software that turns written text into spoken words, so you can listen while you read along.
speech-to-text
Software that turns speech into written text. Dictation is one kind.
summary
A shorter text that leaves some of the original out.
outline
A text cut into headings and short lines under them.
word error rate
The count of wrong, missing and extra words in a transcript, divided by the number of words really said.
Where next
A free research paper about summaries that add things the text did not say. It is written for researchers. It is one page on the ACL Anthology site. In our test on 9 October 2026, it loaded files from four other hosts. They were cdn.jsdelivr.net, use.fontawesome.com, i.creativecommons.org and licensebuttons.net. No cookie was set.
