Simple, you just check whether the header, data structure, metadata, checksum, etc. in the trial-decrypted data in valid, no hack is needed. For example, something as simple as a 16-byte magic number can ensure a false positive rate of 1/2^128. Since any arbitrary binary data can be encrypted, "this is mostly ASCII" is unusable.

In GPG:

    /* if KeyID is empty... */
    if (!k->keyid[0] && !k->keyid[1])
    {
        log_info (_("anonymous recipient; trying secret key %s ...\n"), keystr (keyid));
    }

    err = get_it (ctrl, k, dek, sk, keyid);
    k->result = err;
    if (!err)
    {
        /* If get_it() succeeds */
        if (!opt.quiet && !k->keyid[0] && !k->keyid[1])
        {
            log_info (_("okay, we are the anonymous recipient.\n"));
        }
    }

And in get_it()...

    if (sk->pubkey_algo == PUBKEY_ALGO_ECDH)
    {
        /* Now the frame are the bytes decrypted but padded session key.  */
        if (!nframe || nframe <= 8
                || frame[nframe-1] > nframe)
        {
            err = gpg_error (GPG_ERR_WRONG_SECKEY);
            goto leave;
        }
    }
    else
    {
        if (padding)
        {
            if (n + 7 > nframe)
            {
                err = gpg_error (GPG_ERR_WRONG_SECKEY);
                goto leave;
            }
        }
    
        if (n + 4 > nframe)
        {
            err = gpg_error (GPG_ERR_WRONG_SECKEY);
            goto leave;
        }
    
        if (dek->keylen != openpgp_cipher_get_algo_keylen (dek->algo))
        {
            err = gpg_error (GPG_ERR_WRONG_SECKEY);
            goto leave;
        }
    
        /* Copy the key to DEK and compare the checksum.  */
        csum = buf16_to_u16 (frame+nframe-2);
        memcpy (dek->key, frame + n, dek->keylen);
        for (csum2 = 0, n = 0; n < dek->keylen; n++)
            csum2 += dek->key[n];
        if (csum != csum2)
        {
            err = gpg_error (GPG_ERR_WRONG_SECKEY);
            goto leave;
        }
    }

You get the idea.

BTW, did I just "break GPG encryption and deanonymize users" according to Cellebrite because I found the correct function in the code? And no, I didn't "review dozens of code classes", I just grepped the word "anonymous"... /s