How to interpret OpenAI AI Text Classifier Classes

Question

OpenAI has recently announced a new AI classifier for indicating AI-written text, that is available as a tool here. According to the Model Card this classifier categorize five different classes as follows:

• "Very unlikely to be AI-generated" corresponds to a classifier threshold of <0.1.
• "Unlikely to be AI-generated" corresponds to a classifier threshold between 0.1 and 0.45.
• "Unclear if it is AI written" corresponds to a classifier threshold between 0.45 and 0.9.
• "Possibly AI-generated" corresponds to a classifier threshold between 0.9 and 0.98.
• "Likely AI-generated" corresponds to a classifier threshold >0.98.

But when I try it a get in the output the logprobs of specific tokens that seems to map the classes above. This is an example of log probs in the case of ai-generated text:

{
  "top_logprobs": [
          {
            "!": -3.4727178,
            "\"": -0.033285,
            " \"": -8.678962,
            "\".": -10.159605,
            "!\"": -8.384942
          }
   ]
}

while this is an example of log probs in the case of human-generated text:

{
    "top_logprobs": [
          {
            "!": -0.0272431,
            "\"": -3.7918115,
            "!\n\n": -8.135782,
            "!!": -8.312574,
            "”": -8.102049
          }
     ]
}

to read these easily we could eventually convert log probs into probs doing like

log_probs = response.choices[0].logprobs.top_logprobs[0];
probs = Object.keys(log_probs).map(v => ({ label: v, log_prob: log_probs[v], prob: Math.exp(log_probs[v]) }));
probs = probs.sort((a,b) => (b.prob - a.prob));

Hence getting for AI-Generated Text

[
  {
    "label": "\"",
    "log_prob": -0.033285,
    "prob": 0.9672628503899762
  },
  {
    "label": "!",
    "log_prob": -3.4727178,
    "prob": 0.03103257561308009
  },
  {
    "label": "!\"",
    "log_prob": -8.384942,
    "prob": 0.0002282789983770749
  },
  {
    "label": " \"",
    "log_prob": -8.678962,
    "prob": 0.00017012756839526099
  },
  {
    "label": "\".",
    "log_prob": -10.159605,
    "prob": 0.00003870255265601741
  }
]

and for Human-generated text:

[
  {
    "label": "!",
    "log_prob": -0.0272431,
    "prob": 0.9731246461658533
  },
  {
    "label": "\"",
    "log_prob": -3.7918115,
    "prob": 0.022554706970090203
  },
  {
    "label": "”",
    "log_prob": -8.102049,
    "prob": 0.00030291782313947616
  },
  {
    "label": "!\n\n",
    "log_prob": -8.135782,
    "prob": 0.000292869921924322
  },
  {
    "label": "!!",
    "log_prob": -8.312574,
    "prob": 0.00024541154037504843
  }
]

so it seems that

• the token "!" is related to the classes "Unlikely to be AI-generated" (>0.1 and <0.45) and "Very unlikely to be AI-generated" (<0.1)
• the token "\"" seems to be related to "Likely AI-generated" (>0.98) and "Possibly AI-generated" (>0.9 and <0.98)

For all other probability values ( >0.45 and <0.9 ) the label would be "Unclear if it is AI written".

So to get the label and its prob I did this function where I filter out all tokens but "!" and "\"", turn log probs into prob, sort and assign the label description:

let getProbabilities = function(log_probs) {
        var probs = Object.keys(log_probs).filter(v => v==="!" || v==="\"").map(v => ({ label: v, log_prob: log_probs[v], prob: Math.exp(log_probs[v]) }));
        probs = probs.sort((a,b) => (b.prob - a.prob));
        var predicted = probs[0].label;
        probs = probs.map(pred => {
            if(pred.label==="!" && predicted==pred.label) {  // Human-generated
                if((1-pred.prob)>0.1 && (1-pred.prob)<0.45) pred.desc='Unlikely to be AI-generated'
                else if((1-pred.prob)<0.1) pred.desc='Very unlikely to be AI-generated'
            }
            else if(pred.label==="\"" && predicted==pred.label) { // AI-generated
                if(pred.prob>0.98) pred.desc='Likely AI-generated'
                else if(pred.prob>0.9 && pred.prob<0.98) pred.desc='Possibly AI-generated'
                else if(pred.prob>0.45 && pred.prob<0.9) pred.desc='Unclear if it is AI written'
            }
            return pred;
        });
        return probs;
}//getProbabilities

so I get

log_probs = res.choices[0].logprobs.top_logprobs[0];
probs = getProbabilities(log_probs);
[
  {
    "label": "\"",
    "log_prob": -0.033285,
    "prob": 0.9672628503899762,
    "desc": "Possibly AI-generated"
  },
  {
    "label": "!",
    "log_prob": -3.4727178,
    "prob": 0.03103257561308009
  }
]

and

log_probs = res.choices[0].logprobs.top_logprobs[0];
probs = getProbabilities(log_probs);
[
  {
    "label": "!",
    "log_prob": -0.0272431,
    "prob": 0.9731246461658533,
    "desc": "Very unlikely to be AI-generated"
  },
  {
    "label": "\"",
    "log_prob": -3.7918115,
    "prob": 0.022554706970090203
  }
]

I'm not sure if the assumptions I did in getProbabilities are correct, is out there any better way?