As I understand, Apple is introducing three types of writing tools:
Rewrite, which generates alternative wording in one of three styles: Friendly, Professional, or Concise
Proofread and Summarize, and
Smart Reply
I’m most interested in Proofread’s capabilities but I haven’t found many specifics on exactly what it will do and this is the only screenshot I’m aware of:
Let’s see what Apple will come up with, the best existing tool of this kind I’m aware of is deepl.com. It has four styles
Simple
Business
Academic
Casual
and can set four tones
Enthusiastic
Friendly
Confident
Diplomatic
DeepL’s translation quality is far superior to Google Translate. They also have a free Mac App. The free tier of DeepL has generous allowances. I use it a lot.
Apparently the developer betas for iOS 18.1, iPadOS 18.1, and macOS Sequoia 15.1 released yesterday (July 29) include the proofreading Writing Tool:
With Writing Tools, you can rewrite, spellcheck, grammar check, and summarize text in Messages, Notes, Mail, Pages, and other apps where you write. Summaries also work for transcripts in the Notes app.
Source
Hopefully, someone here on TidBITS Talk will take proofreading for a spin and post what they think of it. I’m particularly interested in using grammar checking in Spark email for iOS and Gmail within Chrome for macOS.
They followed up with another article with some examples for proofreading et al. Some of it looks real nice IMHO.
I was a bit put off though, that after all the AI hoopla of the last years and Apple finally also dancing to the music (or at least to the marketing), one example for the brand new Apple “Intelligence” was a suggestion to answer these questions raised in an email
Please let me know if your partner would like to join, and whether you will drive or take an uber.
with the AI generated response
Let me know if my partner can join too.
Note that this response was generated AFTER the user had already selected Yes for Smart Reply’s correctly posed query “Will your partner be joining?”.
That whole AI sequence to me looks like much more A than I.
Although, granted, it is very human, as in certain current-day humans who have unlearned how to read two sentences and understand them before jumping to reply with garbage. So perhaps this AI just attempts to be more like certain modern day humans than an actually intelligent or supportive assistant.
Joking aside, I hope that’s just bugs early in the process and that by the time this goes live, it’s not full of obviously dumb mistakes like that. In order for this stuff to be broadly adopted it needs to be reliable and efficient — at least for these very simple yes/no queries. If it were to remain hit and miss, why bother taking chances, might as well just type up a quick 3-word reply yourself.
As a professional writer, what I could like to see in the way of computer proofreading (AI is justifiably getting a bad name) is software that would highlight a questionable word, phrase or spelling and ask “is this what you meant to write?” That way I could look at it for myself and make the change (if any) that I want, rather than leaving it up to the software to make a change that I might not spot in checking proofs.
I had a very bad experience when spellchecking software took it upon itself to change technical terms throughout a whole book and I didn’t discover the problem until I was printing the final manuscript (this was in the 1980s, when publishers wanted printed manuscripts). This was a book of 350 pages on fiber optics and communication, and it changed the capitalization of abbreviations to match the capitalization of a name. So MHz (megahertz) and GHz (gigahertz) were changed to Mhz and Ghz, and there were other similar abbreviations through the book.
Unfortunately, it doesn’t include a demonstration of the proofreading capabilities of the Writing Tool, which is the portion of Apple Intelligence that I’m interested in assessing. (If I missed the proofreading portion, please let me know where it starts.)
Just add the AI proofreading part to what is now the spellcheck dotted red underline. Make it blue if necessary to distinguish the two. That allows alerting me to a potential issue along with perhaps making a suggestion for improvement (the same way spellcheck does), but leave it to me to implement any changes.
I would be cautious about suggestions for improvement from any AI. That seems to be a weak point, judging from my experience with the AI feature on Ancestry’s genealogy data. I tried it several times and every time there were subtle errors. It mistook my mother’s grandmother for my mother because my mother had been given was given her grandmother’s first name as her middle name. The AI could not recognize where one short newspaper obituary ended and the next started, scrambling names. These were subtle mistakes that I could catch because they were for relatives I knew, and I decided the AI results were so bad they weren’t worth the effort because I had no way to check them for more distant relatives. I would worry about similar errors for fact-checking with Apple Intelligence, and would not use any suggestions (such as names) without verifying them with other sources.
It seems to me that any such AI “support” only warrants serious consideration under the assumption that, by the time it ships in macOS/iOS, it’s good enough to be useful most of the time. It can make a bad suggestion every once in a while, but most of the time it has to suggest stuff that’s either worthy of straight up acceptance or at least gets you to come up with better wording than if you had had no hint at all. As I mentioned above, if it were to remain hit and miss (or worse), why even bother? Turn it off, do it yourself just like before, and save time while getting rid of clutter. It has to be so good and so reliable, that you overall feel more efficient. It’s quite annoying to have to constantly decline supposedly intelligent suggestions, if not outright aggravating. Although that’s usually reserved for when you have to back out automatically inserted “improvements” (looking at you, Clippy).
This is almost exactly how Grammarly works now. I have macOS 15.1 installing on my M1 MacBook Air as I write, so I’ll be able to compare personally soon.
This recent MacRumors article on Writing Tools includes details about the Proofreading capabilities:
I’m disappointed to learn that:
Unfortunately, Writing Tools does not go through issues one by one with you for making changes on an error-by-error basis. What it does is spit out a rewritten version of what you’ve asked it to proofread, and you can choose to copy that text, replace your text with it, or share it. It’s an all-or-nothing correction, and it can be hard to spot every change if you’re correcting a lot of text.
If that’s actually the case, it would make Writing Tools completely unusable for me as a professional writer. If that’s what “Apple Intelligence” is, it’s really Apple Stupidity. I am dismayed.
I also was hoping for something more interactive, but this is the workflow I already use with ChatGPT and it’s not too bad. I save both versions and use BBEdit’s Find Differences feature to go through each of the changes. It works quite well.
I’m more curious about the changes Apple Intelligence will suggest. ChatGPT’s are not perfect and sometimes detrimental.
(I’m also often surprised and impressed at some of the things it catches. For example, in programming articles I will tag certain programming language keywords as code and in one article I missed one and ChatGPT caught it and tagged it in the correction!)
So far most of my experience with AI has been so bad that I usually avoid it. Ancestry uses it for genealogy and it makes stupid mistakes that are hard to spot if you don’t know the family. For example, when it goes through a series of short newspaper obits, the AI can’t tell where one stops and the next starts, so they pick up a surname from one obit and the given name from another of an unrelated person. I was able to identify the problem, but only because I knew the family well. Siteground has been using AI to tell you how to do something, and the procedures don’t work. The only thing I have found useful is Otter.ai which I use to transcribe interviews. I run the recording through the AI, then listen to the recording and compare it to the transcript. It saves a lot of time and comparing the audio to the transcript catches errors, which happen.
