I wanted to share my experience with Turing on a GDM Personalization / AI Quality Analyst project, because honestly it has been one of the most chaotic onboarding processes I have ever seen.
From day one, everything felt improvised.
People were added to Slack channels without clear instructions, some had access to the tools, others did not, some could see the models, others could not, some had Google Groups access, others were stuck for days.
The main thing everyone needed, the so-called Sian Profile / personalization setup, worked randomly depending on the person, the account, the model, the browser, or whatever internal switch someone forgot to enable.
The whole onboarding was a mess:
- Multiple Critical Verification forms.
- Changing model names.
- Conflicting instructions about whether to upload debug TXT files, full HTML files, or Google Drive links.
- People being told different things by different leads.
- Repeated meetings with very short notice.
- Slack channels multiplying everywhere, but nobody giving one clear official answer.
- Some people being allowed to start tasks while others were still blocked by access issues.
Then there is the payment side, which is even worse.
Some people were hired directly through Turing, others through third-party vendors, and the hourly rates were completely different.
People were talking about $20/hour, $15/hour, even around $11.50/hour depending on the vendor. On top of that, Jibble time tracking was a complete mess.
Some people were told to log hours, others were told to delete them. Some had the right Jibble project/activity, others did not. Some were in the Jibble Slack group, others were not.
Later they introduced an AHT model, where you are paid based on the “approved handling time” for each task. In practice, if you complete a task faster, you log less time and earn less; if it takes longer, your pay is capped unless extra time is approved.
So the incentive becomes absurd: work fast and get paid less, work carefully and risk going over the cap.
The tooling itself was also broken. The labeling tool and autoreviewer constantly failed for ridiculous reasons:
- HTML links being read as binary files.
- Prompt match failing because of hidden JavaScript data.
- The reviewer detecting the wrong first prompt inside the HTML.
- Scores changing after running the same review again without modifying anything.
- Tasks failing because of whitespace or formatting issues.
- Confusing instructions about whether to paste raw HTML or a Drive link.
At one point, production was paused because “1P Sources” were appearing in debug information.
From what we could see, internal/private information was being exposed in the debug output.
That is a serious privacy and security concern, not just a minor workflow issue. The batch was paused, people were told to stop claiming tasks, and many of us were left wondering whether the work already done would be paid.
The review process was another nightmare.
Reviewers started sending tasks back as rework, sometimes based on criteria that had not been clearly explained.
For example, people were told during training to use example prompts as inspiration, then later got penalized for using prompts too similar to the examples.
Some reworks seemed reasonable, but others felt completely inconsistent with the instructions we had been given.
And now, to make it even more absurd, I am receiving emails in my personal inbox containing information related to other people in onboarding. That should not be happening. It adds to the feeling that the whole operation is being held together with duct tape.
The worst part is not even one single issue. It is the combination of everything:
- Chaotic onboarding.
- Broken access management.
- Unclear payment rules.
- Inconsistent communication.
- Privacy concerns.
- Buggy internal tools.
- Repeated verification loops.
- Rework criteria changing after the fact.
- Personal emails receiving onboarding-related information about other people.
I understand that large AI data projects can be messy, especially when they scale quickly. But this felt beyond normal startup chaos. It felt like workers were being onboarded before the process, tooling, access control and payment structure were actually ready.
If anyone is considering working with Turing or through a vendor on one of these AI evaluation projects, document everything. Keep screenshots, task IDs, hours worked, Slack messages, emails, Jibble records and any payment-related communication.
Do not rely on verbal instructions or scattered Slack replies.
This has been a complete organizational disaster.
PD: share your links for other jobs, as this is too much of a mess